Merge pull request #27 from jamsilva/master

Improved sound

12 files changed, 546 insertions, 1064 deletions

diff --git a/source/apu.c b/source/apu.c
index 500a884..efbd4b5 100644
--- a/source/apu.c
+++ b/source/apu.c
@@ -8,7 +8,9 @@
 #include "soundux.h"
 #include "cpuexec.h"
 
-extern int32_t NoiseFreq [32];
+extern int32_t  NoiseFreq       [32];
+extern uint16_t DecreaseRateExp [32];
+extern uint16_t IncreaseRate    [32];
 
 bool S9xInitAPU()
 {
@@ -20,8 +22,6 @@ bool S9xInitAPU()
       return (false);
    }
 
-   memset(IAPU.RAM, 0, 0x10000);
-
    return (true);
 }
 
@@ -40,7 +40,7 @@ void S9xResetAPU()
 {
    int i, j;
 
-   Settings.APUEnabled = Settings.NextAPUEnabled;
+   Settings.APUEnabled = true;
 
    memset(IAPU.RAM, 0, 0x100);
    memset(IAPU.RAM + 0x20, 0xFF, 0x20);
@@ -125,13 +125,11 @@ void S9xSetAPUDSP(uint8_t byte)
       else
       {
          S9xSetEchoWriteEnable(!(byte & APU_ECHO_DISABLED));
-
+         so.mute_sound = !!(byte & APU_MUTE);
          SoundData.noise_hertz = NoiseFreq [byte & 0x1f];
          for (i = 0; i < 8; i++)
-         {
             if (SoundData.channels [i].type == SOUND_NOISE)
                S9xSetSoundFrequency(i, SoundData.noise_hertz);
-         }
       }
       break;
    case APU_NON:
@@ -154,36 +152,23 @@ void S9xSetAPUDSP(uint8_t byte)
       break;
    case APU_MVOL_LEFT:
       if (byte != APU.DSP [APU_MVOL_LEFT])
-      {
-         S9xSetMasterVolume((signed char) byte,
-                            (signed char) APU.DSP [APU_MVOL_RIGHT]);
-      }
+         S9xSetMasterVolume((int8_t) byte, (int8_t) APU.DSP [APU_MVOL_RIGHT]);
       break;
    case APU_MVOL_RIGHT:
       if (byte != APU.DSP [APU_MVOL_RIGHT])
-      {
-         S9xSetMasterVolume((signed char) APU.DSP [APU_MVOL_LEFT],
-                            (signed char) byte);
-      }
+         S9xSetMasterVolume((int8_t) APU.DSP [APU_MVOL_LEFT], (int8_t) byte);
       break;
    case APU_EVOL_LEFT:
       if (byte != APU.DSP [APU_EVOL_LEFT])
-      {
-         S9xSetEchoVolume((signed char) byte,
-                          (signed char) APU.DSP [APU_EVOL_RIGHT]);
-      }
+         S9xSetEchoVolume((int8_t) byte, (int8_t) APU.DSP [APU_EVOL_RIGHT]);
       break;
    case APU_EVOL_RIGHT:
       if (byte != APU.DSP [APU_EVOL_RIGHT])
-      {
-         S9xSetEchoVolume((signed char) APU.DSP [APU_EVOL_LEFT],
-                          (signed char) byte);
-      }
+         S9xSetEchoVolume((int8_t) APU.DSP [APU_EVOL_LEFT], (int8_t) byte);
       break;
    case APU_ENDX:
       byte = 0;
       break;
-
    case APU_KOFF:
    {
       int c;
@@ -194,12 +179,10 @@ void S9xSetAPUDSP(uint8_t byte)
          {
             if (APU.KeyedChannels & mask)
             {
-               {
-                  KeyOnPrev &= ~mask;
-                  APU.KeyedChannels &= ~mask;
-                  APU.DSP [APU_KON] &= ~mask;
-                  S9xSetSoundKeyOff(c);
-               }
+               KeyOnPrev &= ~mask;
+               APU.KeyedChannels &= ~mask;
+               APU.DSP [APU_KON] &= ~mask;
+               S9xSetSoundKeyOff(c);
             }
          }
          else if ((KeyOnPrev & mask) != 0)
@@ -211,9 +194,10 @@ void S9xSetAPUDSP(uint8_t byte)
             S9xPlaySample(c);
          }
       }
+
+      APU.DSP [APU_KOFF] = byte;
+      return;
    }
-   APU.DSP [APU_KOFF] = byte;
-   return;
    case APU_KON:
       if (byte)
       {
@@ -237,7 +221,6 @@ void S9xSetAPUDSP(uint8_t byte)
          }
       }
       return;
-
    case APU_VOL_LEFT + 0x00:
    case APU_VOL_LEFT + 0x10:
    case APU_VOL_LEFT + 0x20:
@@ -246,11 +229,8 @@ void S9xSetAPUDSP(uint8_t byte)
    case APU_VOL_LEFT + 0x50:
    case APU_VOL_LEFT + 0x60:
    case APU_VOL_LEFT + 0x70:
-   {
-      S9xSetSoundVolume(reg >> 4, (signed char) byte,
-                        (signed char) APU.DSP [reg + 1]);
-   }
-   break;
+      S9xSetSoundVolume(reg >> 4, (int8_t) byte, (int8_t) APU.DSP [reg + 1]);
+      break;
    case APU_VOL_RIGHT + 0x00:
    case APU_VOL_RIGHT + 0x10:
    case APU_VOL_RIGHT + 0x20:
@@ -259,12 +239,8 @@ void S9xSetAPUDSP(uint8_t byte)
    case APU_VOL_RIGHT + 0x50:
    case APU_VOL_RIGHT + 0x60:
    case APU_VOL_RIGHT + 0x70:
-   {
-      S9xSetSoundVolume(reg >> 4, (signed char) APU.DSP [reg - 1],
-                        (signed char) byte);
-   }
-   break;
-
+      S9xSetSoundVolume(reg >> 4, (int8_t) APU.DSP [reg - 1], (int8_t) byte);
+      break;
    case APU_P_LOW + 0x00:
    case APU_P_LOW + 0x10:
    case APU_P_LOW + 0x20:
@@ -273,10 +249,8 @@ void S9xSetAPUDSP(uint8_t byte)
    case APU_P_LOW + 0x50:
    case APU_P_LOW + 0x60:
    case APU_P_LOW + 0x70:
-      S9xSetSoundHertz(reg >> 4,
-                       (((byte + (APU.DSP [reg + 1] << 8)) & FREQUENCY_MASK) * 32000) >> 12);
+      S9xSetSoundHertz(reg >> 4, ((byte + (APU.DSP [reg + 1] << 8)) & FREQUENCY_MASK) / 8);
       break;
-
    case APU_P_HIGH + 0x00:
    case APU_P_HIGH + 0x10:
    case APU_P_HIGH + 0x20:
@@ -285,20 +259,8 @@ void S9xSetAPUDSP(uint8_t byte)
    case APU_P_HIGH + 0x50:
    case APU_P_HIGH + 0x60:
    case APU_P_HIGH + 0x70:
-      S9xSetSoundHertz(reg >> 4,
-                       (((byte << 8) + APU.DSP [reg - 1]) & FREQUENCY_MASK) * 8);
-      break;
-
-   case APU_SRCN + 0x00:
-   case APU_SRCN + 0x10:
-   case APU_SRCN + 0x20:
-   case APU_SRCN + 0x30:
-   case APU_SRCN + 0x40:
-   case APU_SRCN + 0x50:
-   case APU_SRCN + 0x60:
-   case APU_SRCN + 0x70:
+      S9xSetSoundHertz(reg >> 4, (((byte << 8) + APU.DSP [reg - 1]) & FREQUENCY_MASK) * 8);
       break;
-
    case APU_ADSR1 + 0x00:
    case APU_ADSR1 + 0x10:
    case APU_ADSR1 + 0x20:
@@ -307,15 +269,9 @@ void S9xSetAPUDSP(uint8_t byte)
    case APU_ADSR1 + 0x50:
    case APU_ADSR1 + 0x60:
    case APU_ADSR1 + 0x70:
-      if (byte != APU.DSP [reg])
-      {
-         {
-            S9xFixEnvelope(reg >> 4, APU.DSP [reg + 2], byte,
-                           APU.DSP [reg + 1]);
-         }
-      }
+      if(byte != APU.DSP [reg])
+         S9xFixEnvelope(reg >> 4, APU.DSP [reg + 2], byte, APU.DSP [reg + 1]);
       break;
-
    case APU_ADSR2 + 0x00:
    case APU_ADSR2 + 0x10:
    case APU_ADSR2 + 0x20:
@@ -324,15 +280,9 @@ void S9xSetAPUDSP(uint8_t byte)
    case APU_ADSR2 + 0x50:
    case APU_ADSR2 + 0x60:
    case APU_ADSR2 + 0x70:
-      if (byte != APU.DSP [reg])
-      {
-         {
-            S9xFixEnvelope(reg >> 4, APU.DSP [reg + 1], APU.DSP [reg - 1],
-                           byte);
-         }
-      }
+      if(byte != APU.DSP [reg])
+         S9xFixEnvelope(reg >> 4, APU.DSP [reg + 1], APU.DSP [reg - 1], byte);
       break;
-
    case APU_GAIN + 0x00:
    case APU_GAIN + 0x10:
    case APU_GAIN + 0x20:
@@ -341,59 +291,23 @@ void S9xSetAPUDSP(uint8_t byte)
    case APU_GAIN + 0x50:
    case APU_GAIN + 0x60:
    case APU_GAIN + 0x70:
-      if (byte != APU.DSP [reg])
-      {
-         {
-            S9xFixEnvelope(reg >> 4, byte, APU.DSP [reg - 2],
-                           APU.DSP [reg - 1]);
-         }
-      }
-      break;
-
-   case APU_ENVX + 0x00:
-   case APU_ENVX + 0x10:
-   case APU_ENVX + 0x20:
-   case APU_ENVX + 0x30:
-   case APU_ENVX + 0x40:
-   case APU_ENVX + 0x50:
-   case APU_ENVX + 0x60:
-   case APU_ENVX + 0x70:
-      break;
-
-   case APU_OUTX + 0x00:
-   case APU_OUTX + 0x10:
-   case APU_OUTX + 0x20:
-   case APU_OUTX + 0x30:
-   case APU_OUTX + 0x40:
-   case APU_OUTX + 0x50:
-   case APU_OUTX + 0x60:
-   case APU_OUTX + 0x70:
+      if(byte != APU.DSP [reg])
+         S9xFixEnvelope(reg >> 4, byte, APU.DSP [reg - 2], APU.DSP [reg - 1]);
       break;
-
-   case APU_DIR:
-      break;
-
    case APU_PMON:
-      if (byte != APU.DSP [APU_PMON])
+      if(byte != APU.DSP [APU_PMON])
          S9xSetFrequencyModulationEnable(byte);
       break;
-
    case APU_EON:
-      if (byte != APU.DSP [APU_EON])
+      if(byte != APU.DSP [APU_EON])
          S9xSetEchoEnable(byte);
       break;
-
    case APU_EFB:
-      S9xSetEchoFeedback((signed char) byte);
-      break;
-
-   case APU_ESA:
+      S9xSetEchoFeedback((int8_t) byte);
       break;
-
    case APU_EDL:
       S9xSetEchoDelay(byte & 0xf);
       break;
-
    case APU_C0:
    case APU_C1:
    case APU_C2:
@@ -402,7 +316,7 @@ void S9xSetAPUDSP(uint8_t byte)
    case APU_C5:
    case APU_C6:
    case APU_C7:
-      S9xSetFilterCoefficient(reg >> 4, (signed char) byte);
+      S9xSetFilterCoefficient(reg >> 4, (int8_t) byte);
       break;
    default:
       break;
@@ -417,93 +331,39 @@ void S9xSetAPUDSP(uint8_t byte)
 
 void S9xFixEnvelope(int channel, uint8_t gain, uint8_t adsr1, uint8_t adsr2)
 {
-   if (adsr1 & 0x80)
+   if (adsr1 & 0x80) // ADSR mode
    {
-      // ADSR mode
-      static unsigned long AttackRate [16] =
-      {
-         4100u, 2600u, 1500u, 1000u, 640u, 380u, 260u, 160u,
-         96u, 64u, 40u, 24u, 16u, 10u, 6u, 1u
-      };
-      static unsigned long DecayRate [8] =
-      {
-         1200u, 740u, 440u, 290u, 180u, 110u, 74u, 37u
-      };
-      static unsigned long SustainRate [32] =
-      {
-         ~0u, 38000u, 28000u, 24000u, 19000u, 14000u, 12000u, 9400u,
-         7100u, 5900u, 4700u, 3500u, 2900u, 2400u, 1800u, 1500u,
-         1200u, 880u, 740u, 590u, 440u, 370u, 290u, 220u,
-         180u, 150u, 110u, 92u, 74u, 55u, 37u, 18u
-      };
       // XXX: can DSP be switched to ADSR mode directly from GAIN/INCREASE/
       // DECREASE mode? And if so, what stage of the sequence does it start
       // at?
-      if (S9xSetSoundMode(channel, MODE_ADSR))
+      if(S9xSetSoundMode(channel, MODE_ADSR))
+         S9xSetSoundADSR(channel, adsr1 & 0xf, (adsr1 >> 4) & 7, adsr2 & 0x1f, (adsr2 >> 5) & 7, 8);
+   } // Gain mode
+   else if ((gain & 0x80) == 0)
+   {
+      if (S9xSetSoundMode(channel, MODE_GAIN))
       {
-         // Hack for ROMs that use a very short attack rate, key on a
-         // channel, then switch to decay mode. e.g. Final Fantasy II.
-
-         int attack = AttackRate [adsr1 & 0xf];
-
-         if (attack == 1 && (!Settings.SoundSync))
-            attack = 0;
-
-         S9xSetSoundADSR(channel, attack,
-                         DecayRate [(adsr1 >> 4) & 7],
-                         SustainRate [adsr2 & 0x1f],
-                         (adsr2 >> 5) & 7, 8);
+         S9xSetEnvelopeRate(channel, 0, 0, gain & 0x7f, 0);
+         S9xSetEnvelopeHeight(channel, gain & 0x7f);
       }
    }
+   else if (gain & 0x40)
+   {
+      // Increase mode
+      if(S9xSetSoundMode(channel, (gain & 0x20) ?
+                         MODE_INCREASE_BENT_LINE :
+                         MODE_INCREASE_LINEAR))
+         S9xSetEnvelopeRate(channel, IncreaseRate [gain & 0x1f], 1, 127, (3 << 28) | gain);
+   }
+   else if (gain & 0x20)
+   {
+      if(S9xSetSoundMode(channel, MODE_DECREASE_EXPONENTIAL))
+         S9xSetEnvelopeRate(channel, DecreaseRateExp [gain & 0x1f] / 2, -1, 0, (4 << 28) | gain);
+   }
    else
    {
-      // Gain mode
-      if ((gain & 0x80) == 0)
-      {
-         if (S9xSetSoundMode(channel, MODE_GAIN))
-         {
-            S9xSetEnvelopeRate(channel, 0, 0, gain & 0x7f);
-            S9xSetEnvelopeHeight(channel, gain & 0x7f);
-         }
-      }
-      else
-      {
-         static unsigned long IncreaseRate [32] =
-         {
-            ~0u, 4100u, 3100u, 2600u, 2000u, 1500u, 1300u, 1000u,
-            770u, 640u, 510u, 380u, 320u, 260u, 190u, 160u,
-            130u, 96u, 80u, 64u, 48u, 40u, 32u, 24u,
-            20u, 16u, 12u, 10u, 8u, 6u, 4u, 2u
-         };
-         static unsigned long DecreaseRateExp [32] =
-         {
-            ~0u, 38000u, 28000u, 24000u, 19000u, 14000u, 12000u, 9400u,
-            7100u, 5900u, 4700u, 3500u, 2900u, 2400u, 1800u, 1500u,
-            1200u, 880u, 740u, 590u, 440u, 370u, 290u, 220u,
-            180u, 150u, 110u, 92u, 74u, 55u, 37u, 18u
-         };
-         if (gain & 0x40)
-         {
-            // Increase mode
-            if (S9xSetSoundMode(channel, (gain & 0x20) ?
-                                MODE_INCREASE_BENT_LINE :
-                                MODE_INCREASE_LINEAR))
-            {
-               S9xSetEnvelopeRate(channel, IncreaseRate [gain & 0x1f],
-                                  1, 127);
-            }
-         }
-         else
-         {
-            uint32_t rate = (gain & 0x20) ? DecreaseRateExp [gain & 0x1f] / 2 :
-                          IncreaseRate [gain & 0x1f];
-            int mode = (gain & 0x20) ? MODE_DECREASE_EXPONENTIAL
-                       : MODE_DECREASE_LINEAR;
-
-            if (S9xSetSoundMode(channel, mode))
-               S9xSetEnvelopeRate(channel, rate, -1, 0);
-         }
-      }
+      if (S9xSetSoundMode(channel, MODE_DECREASE_LINEAR))
+         S9xSetEnvelopeRate(channel, IncreaseRate [gain & 0x1f], -1, 0, (3 << 28) | gain);
    }
 }
 
@@ -550,43 +410,16 @@ void S9xSetAPUControl(uint8_t byte)
          APU.ShowROM = true;
       }
    }
-   else
+   else if (APU.ShowROM)
    {
-      if (APU.ShowROM)
-      {
-         APU.ShowROM = false;
-         // memmove converted: Different mallocs [Neb]
-         // DS2 DMA notes: The APU ROM is not 32-byte aligned [Neb]
-         memcpy(&IAPU.RAM [0xffc0], APU.ExtraRAM, sizeof(APUROM));
-      }
+      APU.ShowROM = false;
+      // memmove converted: Different mallocs [Neb]
+      // DS2 DMA notes: The APU ROM is not 32-byte aligned [Neb]
+      memcpy(&IAPU.RAM [0xffc0], APU.ExtraRAM, sizeof(APUROM));
    }
    IAPU.RAM [0xf1] = byte;
 }
 
-void S9xSetAPUTimer(uint16_t Address, uint8_t byte)
-{
-   IAPU.RAM [Address] = byte;
-
-   switch (Address)
-   {
-   case 0xfa:
-      if ((APU.TimerTarget [0] = IAPU.RAM [0xfa]) == 0)
-         APU.TimerTarget [0] = 0x100;
-      APU.TimerValueWritten [0] = true;
-      break;
-   case 0xfb:
-      if ((APU.TimerTarget [1] = IAPU.RAM [0xfb]) == 0)
-         APU.TimerTarget [1] = 0x100;
-      APU.TimerValueWritten [1] = true;
-      break;
-   case 0xfc:
-      if ((APU.TimerTarget [2] = IAPU.RAM [0xfc]) == 0)
-         APU.TimerTarget [2] = 0x100;
-      APU.TimerValueWritten [2] = true;
-      break;
-   }
-}
-
 uint8_t S9xGetAPUDSP()
 {
    uint8_t reg = IAPU.RAM [0xf2] & 0x7f;
@@ -594,10 +427,6 @@ uint8_t S9xGetAPUDSP()
 
    switch (reg)
    {
-   case APU_KON:
-      break;
-   case APU_KOFF:
-      break;
    case APU_OUTX + 0x00:
    case APU_OUTX + 0x10:
    case APU_OUTX + 0x20:
@@ -606,11 +435,10 @@ uint8_t S9xGetAPUDSP()
    case APU_OUTX + 0x50:
    case APU_OUTX + 0x60:
    case APU_OUTX + 0x70:
-      if (SoundData.channels [reg >> 4].state == SOUND_SILENT)
+      if(SoundData.channels [reg >> 4].state == SOUND_SILENT)
          return (0);
       return ((SoundData.channels [reg >> 4].sample >> 8) |
               (SoundData.channels [reg >> 4].sample & 0xff));
-
    case APU_ENVX + 0x00:
    case APU_ENVX + 0x10:
    case APU_ENVX + 0x20:
@@ -620,9 +448,6 @@ uint8_t S9xGetAPUDSP()
    case APU_ENVX + 0x60:
    case APU_ENVX + 0x70:
       return 0;
-
-   case APU_ENDX:
-      break;
    default:
       break;
    }
diff --git a/source/apu.h b/source/apu.h
index 3f88346..6b40363 100644
--- a/source/apu.h
+++ b/source/apu.h
@@ -70,12 +70,11 @@ int S9xAPUOPrint(char* buffer, uint16_t Address);
 void S9xSetAPUControl(uint8_t byte);
 void S9xSetAPUDSP(uint8_t byte);
 uint8_t S9xGetAPUDSP();
-void S9xSetAPUTimer(uint16_t Address, uint8_t byte);
 bool S9xInitSound();
 void S9xOpenCloseSoundTracingFile(bool);
 void S9xPrintAPUState();
-extern uint16_t S9xAPUCycles [256];   // Scaled cycle lengths
-extern uint16_t S9xAPUCycleLengths [256];   // Raw data.
+extern uint16_t S9xAPUCycles [256];       // Scaled cycle lengths
+extern uint16_t S9xAPUCycleLengths [256]; // Raw data.
 extern void (*S9xApuOpcodes [256])(void);
 
 #define APU_VOL_LEFT 0x00
diff --git a/source/apumem.h b/source/apumem.h
index 74594d6..8a9f09a 100644
--- a/source/apumem.h
+++ b/source/apumem.h
@@ -67,32 +67,24 @@ INLINE uint8_t S9xAPUGetByte(uint32_t Address)
 {
    Address &= 0xffff;
 
-   if (Address <= 0xff && Address >= 0xf0)
-   {
-      if (Address >= 0xf4 && Address <= 0xf7)
-      {
-#ifdef SPC700_SHUTDOWN
-         IAPU.WaitAddress2 = IAPU.WaitAddress1;
-         IAPU.WaitAddress1 = IAPU.PC;
-#endif
-         return (IAPU.RAM [Address]);
-      }
-      else if (Address == 0xf3)
-         return (S9xGetAPUDSP());
-      if (Address >= 0xfd)
-      {
+   if (Address == 0xf3)
+      return S9xGetAPUDSP();
+
+   bool zero = (Address >= 0xfd && Address <= 0xff);
+   uint8_t t = IAPU.RAM [Address];
+
 #ifdef SPC700_SHUTDOWN
-         IAPU.WaitAddress2 = IAPU.WaitAddress1;
-         IAPU.WaitAddress1 = IAPU.PC;
-#endif
-         uint8_t t = IAPU.RAM [Address];
-         IAPU.RAM [Address] = 0;
-         return (t);
-      }
-      return (IAPU.RAM [Address]);
+   if (zero || (Address >= 0xf4 && Address <= 0xf7))
+   {
+      IAPU.WaitAddress2 = IAPU.WaitAddress1;
+      IAPU.WaitAddress1 = IAPU.PC;
    }
-   else
-      return (IAPU.RAM [Address]);
+#endif
+
+   if(zero)
+      IAPU.RAM [Address] = 0;
+
+   return t;
 }
 
 INLINE void S9xAPUSetByte(uint8_t byte, uint32_t Address)
diff --git a/source/cpuexec.c b/source/cpuexec.c
index b88e6f1..c834adf 100644
--- a/source/cpuexec.c
+++ b/source/cpuexec.c
@@ -46,7 +46,6 @@ void S9xMainLoop_SA1_SFX(void)
    for (;;)
    {
       APU_EXECUTE();
-
       if (CPU.Flags)
       {
          if (CPU.Flags & NMI_FLAG)
@@ -57,7 +56,7 @@ void S9xMainLoop_SA1_SFX(void)
                if (CPU.WaitingForInterrupt)
                {
                   CPU.WaitingForInterrupt = false;
-                  CPU.PC++;
+                  ++CPU.PC;
                }
                S9xOpcode_NMI();
             }
@@ -70,7 +69,7 @@ void S9xMainLoop_SA1_SFX(void)
                if (CPU.WaitingForInterrupt)
                {
                   CPU.WaitingForInterrupt = false;
-                  CPU.PC++;
+                  ++CPU.PC;
                }
                if (CPU.IRQActive && !Settings.DisableIRQ)
                {
@@ -127,7 +126,6 @@ void S9xMainLoop_SA1_NoSFX(void)
    for (;;)
    {
       APU_EXECUTE();
-
       if (CPU.Flags)
       {
          if (CPU.Flags & NMI_FLAG)
@@ -199,7 +197,6 @@ void S9xMainLoop_NoSA1_SFX(void)
    for (;;)
    {
       APU_EXECUTE();
-
       if (CPU.Flags)
       {
          if (CPU.Flags & NMI_FLAG)
@@ -210,7 +207,7 @@ void S9xMainLoop_NoSA1_SFX(void)
                if (CPU.WaitingForInterrupt)
                {
                   CPU.WaitingForInterrupt = false;
-                  CPU.PC++;
+                  ++CPU.PC;
                }
                S9xOpcode_NMI();
             }
@@ -223,7 +220,7 @@ void S9xMainLoop_NoSA1_SFX(void)
                if (CPU.WaitingForInterrupt)
                {
                   CPU.WaitingForInterrupt = false;
-                  CPU.PC++;
+                  ++CPU.PC;
                }
                if (CPU.IRQActive && !Settings.DisableIRQ)
                {
@@ -278,7 +275,6 @@ void S9xMainLoop_NoSA1_NoSFX(void)
    for (;;)
    {
       APU_EXECUTE();
-
       if (CPU.Flags)
       {
          if (CPU.Flags & NMI_FLAG)
@@ -381,7 +377,6 @@ void S9xDoHBlankProcessing_SFX()
    case HBLANK_START_EVENT:
       if (IPPU.HDMA && CPU.V_Counter <= PPU.ScreenHeight)
          IPPU.HDMA = S9xDoHDMA(IPPU.HDMA);
-
       break;
 
    case HBLANK_END_EVENT:
@@ -390,12 +385,7 @@ void S9xDoHBlankProcessing_SFX()
 #ifndef USE_BLARGG_APU
       CPU.Cycles -= Settings.H_Max;
       if (IAPU.APUExecuting)
-      {
          APU.Cycles -= Settings.H_Max;
-#ifdef MK_APU
-         S9xCatchupCount();
-#endif
-      }
       else
          APU.Cycles = 0;
 #else
@@ -455,17 +445,6 @@ void S9xDoHBlankProcessing_SFX()
             CPU.Flags |= NMI_FLAG;
             CPU.NMICycleCount = CPU.NMITriggerPoint;
          }
-
-#ifdef OLD_SNAPSHOT_CODE
-         if (CPU.Flags & SAVE_SNAPSHOT_FLAG)
-         {
-            CPU.Flags &= ~SAVE_SNAPSHOT_FLAG;
-            Registers.PC = CPU.PC - CPU.PCBase;
-            S9xPackStatus();
-            S9xAPUPackStatus();
-            Snapshot(NULL);
-         }
-#endif
       }
 
       if (CPU.V_Counter == PPU.ScreenHeight + 3)
@@ -554,7 +533,6 @@ void S9xDoHBlankProcessing_NoSFX()
    case HBLANK_START_EVENT:
       if (IPPU.HDMA && CPU.V_Counter <= PPU.ScreenHeight)
          IPPU.HDMA = S9xDoHDMA(IPPU.HDMA);
-
       break;
 
    case HBLANK_END_EVENT:
@@ -562,12 +540,7 @@ void S9xDoHBlankProcessing_NoSFX()
 #ifndef USE_BLARGG_APU
       CPU.Cycles -= Settings.H_Max;
       if (IAPU.APUExecuting)
-      {
          APU.Cycles -= Settings.H_Max;
-#ifdef MK_APU
-         S9xCatchupCount();
-#endif
-      }
       else
          APU.Cycles = 0;
 #else
@@ -579,8 +552,7 @@ void S9xDoHBlankProcessing_NoSFX()
       CPU.NextEvent = -1;
       ICPU.Scanline++;
 
-      if (++CPU.V_Counter >= (Settings.PAL ? SNES_MAX_PAL_VCOUNTER :
-                              SNES_MAX_NTSC_VCOUNTER))
+      if (++CPU.V_Counter >= (Settings.PAL ? SNES_MAX_PAL_VCOUNTER : SNES_MAX_NTSC_VCOUNTER))
       {
          CPU.V_Counter = 0;
          Memory.FillRAM[0x213F] ^= 0x80;
@@ -628,17 +600,6 @@ void S9xDoHBlankProcessing_NoSFX()
             CPU.Flags |= NMI_FLAG;
             CPU.NMICycleCount = CPU.NMITriggerPoint;
          }
-
-#ifdef OLD_SNAPSHOT_CODE
-         if (CPU.Flags & SAVE_SNAPSHOT_FLAG)
-         {
-            CPU.Flags &= ~SAVE_SNAPSHOT_FLAG;
-            Registers.PC = CPU.PC - CPU.PCBase;
-            S9xPackStatus();
-            S9xAPUPackStatus();
-            Snapshot(NULL);
-         }
-#endif
       }
 
       if (CPU.V_Counter == PPU.ScreenHeight + 3)
@@ -705,14 +666,11 @@ void S9xDoHBlankProcessing_NoSFX()
       }
 #endif
       break;
-
    case HTIMER_BEFORE_EVENT:
    case HTIMER_AFTER_EVENT:
-      if (PPU.HTimerEnabled && (!PPU.VTimerEnabled
-                                || CPU.V_Counter == PPU.IRQVBeamPos))
+      if (PPU.HTimerEnabled && (!PPU.VTimerEnabled || CPU.V_Counter == PPU.IRQVBeamPos))
          S9xSetIRQ(PPU_H_BEAM_IRQ_SOURCE);
       break;
    }
-
    S9xReschedule();
 }
diff --git a/source/dma.c b/source/dma.c
index fc1356b..f404d4f 100644
--- a/source/dma.c
+++ b/source/dma.c
@@ -31,13 +31,6 @@ extern uint8_t* HDMABasePointers [8];
 uint32_t HDMARawPointers[8]; // Cart address space pointer
 #endif
 
-#if defined(__linux__) || defined(__WIN32__)
-static int S9xCompareSDD1IndexEntries(const void* p1, const void* p2)
-{
-   return (*(uint32_t*) p1 - * (uint32_t*) p2);
-}
-#endif
-
 /**********************************************************************************************/
 /* S9xDoDMA()                                                                                   */
 /* This function preforms the general dma transfer                                            */
diff --git a/source/fxinst.h b/source/fxinst.h
index 7cb0d2e..f7450a9 100644
--- a/source/fxinst.h
+++ b/source/fxinst.h
@@ -280,9 +280,6 @@ struct FxRegs_s
 /* Update pipe from ROM */
 #define FETCHPIPE { PIPE = PRGBANK(R15); }
 
-/* ABS */
-#define ABS(x) ((x)<0?-(x):(x))
-
 /* Access source register */
 #define SREG (*GSU.pvSreg)
 
diff --git a/source/globals.c b/source/globals.c
index 83adf78..e65d1b5 100644
--- a/source/globals.c
+++ b/source/globals.c
@@ -85,7 +85,6 @@ SCheatData Cheat;
 SoundStatus so;
 
 int32_t Echo [24000];
-int32_t DummyEchoBuffer [SOUND_BUFFER_SIZE];
 int32_t MixBuffer [SOUND_BUFFER_SIZE];
 int32_t EchoBuffer [SOUND_BUFFER_SIZE];
 int32_t FilterTaps [8];
diff --git a/source/sa1.c b/source/sa1.c
index ac423f5..17e7258 100644
--- a/source/sa1.c
+++ b/source/sa1.c
@@ -36,8 +36,7 @@ void S9xSA1Init()
 void S9xSA1Reset()
 {
    SA1.Registers.PB = 0;
-   SA1.Registers.PC = Memory.FillRAM [0x2203] |
-                      (Memory.FillRAM [0x2204] << 8);
+   SA1.Registers.PC = Memory.FillRAM [0x2203] | (Memory.FillRAM [0x2204] << 8);
    SA1.Registers.D.W = 0;
    SA1.Registers.DB = 0;
    SA1.Registers.SH = 1;
@@ -135,8 +134,6 @@ uint8_t S9xSA1GetByte(uint32_t address)
          return ((SA1.BWRAM [(address >> 2) & 0xffff] >> ((address & 3) << 1)) & 3);
       else
          return ((SA1.BWRAM [(address >> 1) & 0xffff] >> ((address & 1) << 2)) & 15);
-
-   case MAP_DEBUG:
    default:
       return OpenBus;
    }
@@ -224,47 +221,35 @@ void S9xSA1SetPCBase(uint32_t address)
    {
    case MAP_PPU:
       SA1.PCBase = Memory.FillRAM - 0x2000;
-      SA1.PC = SA1.PCBase + (address & 0xffff);
-      return;
-
+      break;
    case MAP_CPU:
       SA1.PCBase = Memory.FillRAM - 0x4000;
-      SA1.PC = SA1.PCBase + (address & 0xffff);
-      return;
-
+      break;
    case MAP_DSP:
       SA1.PCBase = Memory.FillRAM - 0x6000;
-      SA1.PC = SA1.PCBase + (address & 0xffff);
-      return;
-
+      break;
    case MAP_SA1RAM:
    case MAP_LOROM_SRAM:
       SA1.PCBase = Memory.SRAM;
-      SA1.PC = SA1.PCBase + (address & 0xffff);
-      return;
-
+      break;
    case MAP_BWRAM:
       SA1.PCBase = SA1.BWRAM - 0x6000;
-      SA1.PC = SA1.PCBase + (address & 0xffff);
-      return;
+      break;
    case MAP_HIROM_SRAM:
       SA1.PCBase = Memory.SRAM - 0x6000;
-      SA1.PC = SA1.PCBase + (address & 0xffff);
-      return;
-
-   case MAP_DEBUG:
-
+      break;
    default:
-   case MAP_NONE:
       SA1.PCBase = Memory.RAM;
-      SA1.PC = Memory.RAM + (address & 0xffff);
-      return;
+      break;
    }
+
+   SA1.PC = SA1.PCBase + (address & 0xffff);
 }
 
 void S9xSetSA1MemMap(uint32_t which1, uint8_t map)
 {
    int c;
+   int i;
    int start = which1 * 0x100 + 0xc00;
    int start2 = which1 * 0x200;
 
@@ -274,8 +259,6 @@ void S9xSetSA1MemMap(uint32_t which1, uint8_t map)
    for (c = 0; c < 0x100; c += 16)
    {
       uint8_t* block = &Memory.ROM [(map & 7) * 0x100000 + (c << 12)];
-      int i;
-
       for (i = c; i < c + 16; i++)
          Memory.Map [start + i] = SA1.Map [start + i] = block;
    }
@@ -283,8 +266,6 @@ void S9xSetSA1MemMap(uint32_t which1, uint8_t map)
    for (c = 0; c < 0x200; c += 16)
    {
       uint8_t* block = &Memory.ROM [(map & 7) * 0x100000 + (c << 11) - 0x8000];
-      int i;
-
       for (i = c + 8; i < c + 16; i++)
          Memory.Map [start2 + i] = SA1.Map [start2 + i] = block;
    }
@@ -296,12 +277,12 @@ uint8_t S9xGetSA1(uint32_t address)
    {
    case 0x2300:
       return ((uint8_t)((Memory.FillRAM [0x2209] & 0x5f) |
-                      (CPU.IRQActive & (SA1_IRQ_SOURCE | SA1_DMA_IRQ_SOURCE))));
+                        (CPU.IRQActive & (SA1_IRQ_SOURCE | SA1_DMA_IRQ_SOURCE))));
    case 0x2301:
       return ((Memory.FillRAM [0x2200] & 0xf) |
               (Memory.FillRAM [0x2301] & 0xf0));
    case 0x2306:
-      return ((uint8_t)  SA1.sum);
+      return ((uint8_t) SA1.sum);
    case 0x2307:
       return ((uint8_t)(SA1.sum >>  8));
    case 0x2308:
@@ -310,8 +291,6 @@ uint8_t S9xGetSA1(uint32_t address)
       return ((uint8_t)(SA1.sum >> 24));
    case 0x230a:
       return ((uint8_t)(SA1.sum >> 32));
-   case 0x230c:
-      return (Memory.FillRAM [0x230c]);
    case 0x230d:
    {
       uint8_t byte = Memory.FillRAM [0x230d];
@@ -321,14 +300,15 @@ uint8_t S9xGetSA1(uint32_t address)
       return (byte);
    }
    default:
-      printf("R: %04x\n", address);
-      break;
+      return (Memory.FillRAM [address]);
    }
-   return (Memory.FillRAM [address]);
 }
 
 void S9xSetSA1(uint8_t byte, uint32_t address)
 {
+   if (address < 0x2200 || address > 0x22ff)
+      return;
+
    switch (address)
    {
    case 0x2200:
@@ -347,11 +327,8 @@ void S9xSetSA1(uint8_t byte, uint32_t address)
          }
       }
       if (byte & 0x10)
-      {
          Memory.FillRAM [0x2301] |= 0x10;
-      }
       break;
-
    case 0x2201:
       if (((byte ^ Memory.FillRAM [0x2201]) & 0x80) &&
             (Memory.FillRAM [0x2300] & byte & 0x80))
@@ -372,15 +349,6 @@ void S9xSetSA1(uint8_t byte, uint32_t address)
          S9xClearIRQ(SA1_DMA_IRQ_SOURCE);
       }
       break;
-
-   case 0x2203:
-   case 0x2204:
-   case 0x2205:
-   case 0x2206:
-   case 0x2207:
-   case 0x2208:
-      break;
-
    case 0x2209:
       Memory.FillRAM [0x2209] = byte;
       if (byte & 0x80)
@@ -388,7 +356,7 @@ void S9xSetSA1(uint8_t byte, uint32_t address)
 
       if (byte & Memory.FillRAM [0x2201] & 0x80)
          S9xSetIRQ(SA1_IRQ_SOURCE);
-      break;
+      return;
    case 0x220a:
       if (((byte ^ Memory.FillRAM [0x220a]) & 0x80) &&
             (Memory.FillRAM [0x2301] & byte & 0x80))
@@ -426,35 +394,10 @@ void S9xSetSA1(uint8_t byte, uint32_t address)
          Memory.FillRAM [0x2301] &= ~0x20;
       }
       if (byte & 0x10)
-      {
-         // Clear NMI
-         Memory.FillRAM [0x2301] &= ~0x10;
-      }
+         Memory.FillRAM [0x2301] &= ~0x10; // Clear NMI
       if (!SA1.IRQActive)
          SA1.Flags &= ~IRQ_PENDING_FLAG;
       break;
-   case 0x220c:
-   case 0x220d:
-   case 0x220e:
-   case 0x220f:
-   case 0x2210:
-      break;
-
-   case 0x2211:
-      printf("Timer reset\n");
-      break;
-   case 0x2212:
-      printf("H-Timer %04x\n", byte | (Memory.FillRAM [0x2213] << 8));
-      break;
-   case 0x2213:
-      printf("H-Timer %04x\n", (byte << 8) | Memory.FillRAM [0x2212]);
-      break;
-   case 0x2214:
-      printf("V-Timer %04x\n", byte | (Memory.FillRAM [0x2215] << 8));
-      break;
-   case 0x2215:
-      printf("V-Timer %04x\n", (byte << 8) | Memory.FillRAM [0x2214]);
-      break;
    case 0x2220:
    case 0x2221:
    case 0x2222:
@@ -468,29 +411,15 @@ void S9xSetSA1(uint8_t byte, uint32_t address)
       if (byte != Memory.FillRAM [address])
          S9xSA1SetBWRAMMemMap(byte);
       break;
-   case 0x2226:
-   case 0x2227:
-   case 0x2228:
-   case 0x2229:
-   case 0x222a:
-   case 0x2230:
-      break;
    case 0x2231:
       if (byte & 0x80)
          SA1.in_char_dma = false;
       break;
-   case 0x2232:
-   case 0x2233:
-   case 0x2234:
-   case 0x2235:
-      Memory.FillRAM [address] = byte;
-      break;
    case 0x2236:
       Memory.FillRAM [address] = byte;
       if ((Memory.FillRAM [0x2230] & 0xa4) == 0x80)
       {
-         // Normal DMA to I-RAM
-         S9xSA1DMA();
+         S9xSA1DMA(); // Normal DMA to I-RAM
       }
       else if ((Memory.FillRAM [0x2230] & 0xb0) == 0xb0)
       {
@@ -499,60 +428,33 @@ void S9xSetSA1(uint8_t byte, uint32_t address)
             S9xSetIRQ(SA1_DMA_IRQ_SOURCE);
          SA1.in_char_dma = true;
       }
-      break;
+      return;
    case 0x2237:
       Memory.FillRAM [address] = byte;
       if ((Memory.FillRAM [0x2230] & 0xa4) == 0x84)
-      {
-         // Normal DMA to BW-RAM
-         S9xSA1DMA();
-      }
-      break;
-   case 0x2238:
-   case 0x2239:
-      Memory.FillRAM [address] = byte;
-      break;
+         S9xSA1DMA(); // Normal DMA to BW-RAM
+      return;
    case 0x223f:
       SA1.VirtualBitmapFormat = (byte & 0x80) ? 2 : 4;
       break;
-
-   case 0x2240:
-   case 0x2241:
-   case 0x2242:
-   case 0x2243:
-   case 0x2244:
-   case 0x2245:
-   case 0x2246:
-   case 0x2247:
-   case 0x2248:
-   case 0x2249:
-   case 0x224a:
-   case 0x224b:
-   case 0x224c:
-   case 0x224d:
-   case 0x224e:
-      Memory.FillRAM [address] = byte;
-      break;
-
    case 0x224f:
       Memory.FillRAM [address] = byte;
       if ((Memory.FillRAM [0x2230] & 0xb0) == 0xa0)
       {
          // Char conversion 2 DMA enabled
          // memmove converted: Same malloc but constant non-overlapping addresses [Neb]
-         memcpy(&Memory.ROM [MAX_ROM_SIZE - 0x10000] + SA1.in_char_dma * 16,
+         memcpy(&Memory.ROM [MAX_ROM_SIZE - 0x10000] + (SA1.in_char_dma << 4),
                 &Memory.FillRAM [0x2240], 16);
          SA1.in_char_dma = (SA1.in_char_dma + 1) & 7;
          if ((SA1.in_char_dma & 3) == 0)
             S9xSA1CharConv2();
       }
-      break;
+      return;
    case 0x2250:
       if (byte & 2)
          SA1.sum = 0;
       SA1.arithmetic_op = byte & 3;
       break;
-
    case 0x2251:
       SA1.op1 = (SA1.op1 & 0xff00) | byte;
       break;
@@ -566,7 +468,7 @@ void S9xSetSA1(uint8_t byte, uint32_t address)
       SA1.op2 = (SA1.op2 & 0xff) | (byte << 8);
       switch (SA1.arithmetic_op)
       {
-      case 0:   // multiply
+      case 0: // multiply
          SA1.sum = SA1.op1 * SA1.op2;
          break;
       case 1: // divide
@@ -574,11 +476,10 @@ void S9xSetSA1(uint8_t byte, uint32_t address)
             SA1.sum = SA1.op1 << 16;
          else
          {
-            SA1.sum = (SA1.op1 / (int)((uint16_t) SA1.op2)) |
-                      ((SA1.op1 % (int)((uint16_t) SA1.op2)) << 16);
+            uint64_t x = (SA1.op1 / (uint16_t) SA1.op2);
+            SA1.sum = x | ((SA1.op1 - (x * (uint16_t) SA1.op2)) << 16);
          }
          break;
-      case 2:
       default: // cumulative sum
          SA1.sum += SA1.op1 * SA1.op2;
          if (SA1.sum & ((int64_t) 0xffffff << 32))
@@ -598,16 +499,13 @@ void S9xSetSA1(uint8_t byte, uint32_t address)
       SA1.variable_bit_pos = 0;
       S9xSA1ReadVariableLengthData(false, true);
       return;
-   default:
-      break;
    }
-   if (address >= 0x2200 && address <= 0x22ff)
-      Memory.FillRAM [address] = byte;
+   
+   Memory.FillRAM [address] = byte;
 }
 
 static void S9xSA1CharConv2()
 {
-   int l, b;
    uint32_t dest = Memory.FillRAM [0x2235] | (Memory.FillRAM [0x2236] << 8);
    uint32_t offset = (SA1.in_char_dma & 7) ? 0 : 1;
    int depth = (Memory.FillRAM [0x2231] & 3) == 0 ? 8 :
@@ -616,15 +514,12 @@ static void S9xSA1CharConv2()
    uint8_t* p = &Memory.FillRAM [0x3000] + dest + offset * bytes_per_char;
    uint8_t* q = &Memory.ROM [MAX_ROM_SIZE - 0x10000] + offset * 64;
 
-   switch (depth)
+   if (depth == 8)
    {
-   case 2:
-      break;
-   case 4:
-      break;
-   case 8:
+      int l;
       for (l = 0; l < 8; l++, q += 8)
       {
+         int b;
          for (b = 0; b < 8; b++)
          {
             uint8_t r = *(q + b);
@@ -639,19 +534,18 @@ static void S9xSA1CharConv2()
          }
          p += 2;
       }
-      break;
    }
 }
 
 static void S9xSA1DMA()
 {
-   uint32_t src =  Memory.FillRAM [0x2232] |
+   uint32_t src = Memory.FillRAM [0x2232] |
                  (Memory.FillRAM [0x2233] << 8) |
                  (Memory.FillRAM [0x2234] << 16);
-   uint32_t dst =  Memory.FillRAM [0x2235] |
+   uint32_t dst = Memory.FillRAM [0x2235] |
                  (Memory.FillRAM [0x2236] << 8) |
                  (Memory.FillRAM [0x2237] << 16);
-   uint32_t len =  Memory.FillRAM [0x2238] |
+   uint32_t len = Memory.FillRAM [0x2238] |
                  (Memory.FillRAM [0x2239] << 8);
 
    uint8_t* s;
@@ -672,7 +566,6 @@ static void S9xSA1DMA()
       s = Memory.SRAM + src;
       break;
    default:
-   case 2:
       src &= 0x3ff;
       len &= 0x3ff;
       s = &Memory.FillRAM [0x3000] + src;
@@ -704,7 +597,7 @@ static void S9xSA1DMA()
 
 void S9xSA1ReadVariableLengthData(bool inc, bool no_shift)
 {
-   uint32_t addr =  Memory.FillRAM [0x2259] |
+   uint32_t addr = Memory.FillRAM [0x2259] |
                   (Memory.FillRAM [0x225a] << 8) |
                   (Memory.FillRAM [0x225b] << 16);
    uint8_t shift = Memory.FillRAM [0x2258] & 15;
@@ -722,7 +615,7 @@ void S9xSA1ReadVariableLengthData(bool inc, bool no_shift)
       s &= 15;
    }
    uint32_t data = S9xSA1GetWord(addr) |
-                 (S9xSA1GetWord(addr + 2) << 16);
+                  (S9xSA1GetWord(addr + 2) << 16);
 
    data >>= s;
    Memory.FillRAM [0x230c] = (uint8_t) data;
diff --git a/source/sa1.h b/source/sa1.h
index 0958fc2..079fef5 100644
--- a/source/sa1.h
+++ b/source/sa1.h
@@ -104,7 +104,7 @@ static inline void S9xSA1PackStatus(void)
 {
    SA1.Registers.PL &= ~(Zero | Negative | Carry | Overflow);
    SA1.Registers.PL |= SA1._Carry | ((SA1._Zero == 0) << 1) |
-                       (SA1._Negative & 0x80) | (SA1._Overflow << 6);
+                      (SA1._Negative & 0x80) | (SA1._Overflow << 6);
 }
 
 static inline void S9xSA1FixCycles(void)
diff --git a/source/soundux.c b/source/soundux.c
index 280c9f9..9e71bae 100644
--- a/source/soundux.c
+++ b/source/soundux.c
@@ -6,21 +6,19 @@
 #include <stdio.h>
 #include <string.h>
 #include <errno.h>
-#include <fcntl.h>
 
 #define CLIP16(v) \
-   if ((v) < -32768) \
+if ((v) < -32768) \
     (v) = -32768; \
-   else \
-   if ((v) > 32767) \
-(v) = 32767
+else \
+if ((v) > 32767) \
+    (v) = 32767
 
 #define CLIP8(v) \
-   if ((v) < -128) \
+if ((v) < -128) \
     (v) = -128; \
-   else \
-   if ((v) > 127) \
-(v) = 127
+else if ((v) > 127) \
+         (v) = 127
 
 #include "snes9x.h"
 #include "soundux.h"
@@ -29,7 +27,6 @@
 #include "cpuexec.h"
 
 extern int32_t Echo [24000];
-extern int32_t DummyEchoBuffer [SOUND_BUFFER_SIZE];
 extern int32_t MixBuffer [SOUND_BUFFER_SIZE];
 extern int32_t EchoBuffer [SOUND_BUFFER_SIZE];
 extern int32_t FilterTaps [8];
@@ -40,6 +37,42 @@ extern int32_t Loop [16];
 
 extern long FilterValues[4][2];
 extern int32_t NoiseFreq [32];
+uint32_t AttackRate [16] =
+{
+   4100, 2600, 1500, 1000, 640, 380, 260, 160,
+   96,   64,   40,   24,   16,  10,  6,   1
+};
+
+uint32_t DecayRate [8] =
+{
+   1200, 740, 440, 290, 180, 110, 74, 37
+};
+
+uint32_t DecreaseRateExp [32] =
+{
+   0xFFFFFFFF, 38000, 28000, 24000, 19000, 14000, 12000, 9400,
+   7100,       5900,  4700,  3500,  2900,  2400,  1800,  1500,
+   1200,       880,   740,   590,   440,   370,   290,   220,
+   180,        150,   110,   92,    74,    55,    37,    18
+};
+
+uint32_t IncreaseRate [32] =
+{
+   0xFFFFFFFF, 4100, 3100, 2600, 2000, 1500, 1300, 1000,
+   770,        640,  510,  380,  320,  260,  190,  160,
+   130,        96,   80,   64,   48,   40,   32,   24,
+   20,         16,   12,   10,   8,    6,    4,    2
+};
+
+#define SustainRate DecreaseRateExp
+
+// precalculated env rates for S9xSetEnvRate
+uint32_t AttackERate     [16][10];
+uint32_t DecayERate       [8][10];
+uint32_t SustainERate    [32][10];
+uint32_t IncreaseERate   [32][10];
+uint32_t DecreaseERateExp[32][10];
+uint32_t KeyOffERate         [10];
 
 static int32_t noise_gen;
 
@@ -51,7 +84,6 @@ static int32_t noise_gen;
 #define ENVX_SHIFT 24
 
 void DecodeBlockAsm(int8_t*, int16_t*, int32_t*, int32_t*);
-void DecodeBlockAsm2(int8_t*, int16_t*, int32_t*, int32_t*);
 
 // F is channel's current frequency and M is the 16-bit modulation waveform
 // from the previous channel multiplied by the current envelope volume level.
@@ -60,17 +92,6 @@ void DecodeBlockAsm2(int8_t*, int16_t*, int32_t*, int32_t*);
 #define LAST_SAMPLE 0xffffff
 #define JUST_PLAYED_LAST_SAMPLE(c) ((c)->sample_pointer >= LAST_SAMPLE)
 
-void S9xSetEightBitConsoleSound(bool Enabled)
-{
-   if (Settings.EightBitConsoleSound != Enabled)
-   {
-      Settings.EightBitConsoleSound = Enabled;
-      int i;
-      for (i = 0; i < 8; i++)
-         SoundData.channels[i].needs_decode = true;
-   }
-}
-
 static inline uint8_t* S9xGetSampleAddress(int sample_number)
 {
    uint32_t addr = (((APU.DSP[APU_DIR] << 8) + (sample_number << 2)) & 0xffff);
@@ -92,7 +113,7 @@ void S9xAPUSetEndX(int ch)
    APU.DSP [APU_ENDX] |= 1 << ch;
 }
 
-void S9xSetEnvRate(Channel* ch, uint32_t rate, int direction, int target)
+void S9xSetEnvRate(Channel* ch, uint32_t rate, int direction, int target, unsigned int mode)
 {
    ch->envx_target = target;
 
@@ -104,38 +125,39 @@ void S9xSetEnvRate(Channel* ch, uint32_t rate, int direction, int target)
    else
       ch->direction = direction;
 
-   static int64_t steps [] =
-   {
-      // 0, 64, 1238, 1238, 256, 1, 64, 109, 64, 1238
-      0,
-      (int64_t) FIXED_POINT * 1000 * 64,
-      (int64_t) FIXED_POINT * 1000 * 619,
-      (int64_t) FIXED_POINT * 1000 * 619,
-      (int64_t) FIXED_POINT * 1000 * 128,
-      (int64_t) FIXED_POINT * 1000 * 1,
-      (int64_t) FIXED_POINT * 1000 * 64,
-      (int64_t) FIXED_POINT * 1000 * 55,
-      (int64_t) FIXED_POINT * 1000 * 64,
-      (int64_t) FIXED_POINT * 1000 * 619
-   };
-
    if (rate == 0 || so.playback_rate == 0)
       ch->erate = 0;
    else
    {
-      ch->erate = (uint32_t)
-                  (steps [ch->state] / (rate * so.playback_rate));
+      switch (mode >> 28)
+      {
+      case 0: // Attack
+         ch->erate = AttackERate[ch->env_ind_attack][ch->state];
+         break;
+      case 1: // Decay
+         ch->erate = DecayERate[ch->env_ind_decay][ch->state];
+         break;
+      case 2: // Sustain
+         ch->erate = SustainERate[ch->env_ind_sustain][ch->state];
+         break;
+      case 3: // Increase
+         ch->erate = IncreaseERate[mode & 0x1f][ch->state];
+         break;
+      case 4: // DecreaseExp
+         ch->erate = DecreaseERateExp[mode & 0x1f][ch->state];
+         break;
+      case 5: // KeyOff
+         ch->erate = KeyOffERate[ch->state];
+         break;
+      }
    }
 }
 
-
-void S9xSetEnvelopeRate(int channel, uint32_t rate, int direction,
-                        int target)
+void S9xSetEnvelopeRate(int channel, uint32_t rate, int direction, int target, unsigned int mode)
 {
-   S9xSetEnvRate(&SoundData.channels [channel], rate, direction, target);
+   S9xSetEnvRate(&SoundData.channels [channel], rate, direction, target, mode);
 }
 
-
 void S9xSetSoundVolume(int channel, int16_t volume_left, int16_t volume_right)
 {
    Channel* ch = &SoundData.channels[channel];
@@ -180,7 +202,7 @@ void S9xSetEchoEnable(uint8_t byte)
       if (byte & (1 << i))
          SoundData.channels [i].echo_buf_ptr = EchoBuffer;
       else
-         SoundData.channels [i].echo_buf_ptr = DummyEchoBuffer;
+         SoundData.channels [i].echo_buf_ptr = NULL;
    }
 }
 
@@ -192,7 +214,7 @@ void S9xSetEchoFeedback(int feedback)
 
 void S9xSetEchoDelay(int delay)
 {
-   SoundData.echo_buffer_size = (512 * delay * so.playback_rate) / 32000;
+   SoundData.echo_buffer_size = (512 * delay * so.playback_rate) >> 15;
    SoundData.echo_buffer_size <<= 1;
    if (SoundData.echo_buffer_size)
       SoundData.echo_ptr %= SoundData.echo_buffer_size;
@@ -209,7 +231,7 @@ void S9xSetEchoWriteEnable(uint8_t byte)
 
 void S9xSetFrequencyModulationEnable(uint8_t byte)
 {
-   SoundData.pitch_mod = byte & ~1;
+   SoundData.pitch_mod = byte & 0xFE;
 }
 
 void S9xSetSoundKeyOff(int channel)
@@ -220,7 +242,7 @@ void S9xSetSoundKeyOff(int channel)
    {
       ch->state = SOUND_RELEASE;
       ch->mode = MODE_RELEASE;
-      S9xSetEnvRate(ch, 8, -1, 0);
+      S9xSetEnvRate(ch, 8, -1, 0, 5 << 28);
    }
 }
 
@@ -247,10 +269,8 @@ void S9xFixSoundAfterSnapshotLoad()
       SoundData.channels [i].envxx = SoundData.channels [i].envx << ENVX_SHIFT;
       SoundData.channels [i].next_sample = 0;
       SoundData.channels [i].interpolate = 0;
-      SoundData.channels [i].previous [0] = (int32_t)
-                                            SoundData.channels [i].previous16 [0];
-      SoundData.channels [i].previous [1] = (int32_t)
-                                            SoundData.channels [i].previous16 [1];
+      SoundData.channels [i].previous [0] = (int32_t) SoundData.channels [i].previous16 [0];
+      SoundData.channels [i].previous [1] = (int32_t) SoundData.channels [i].previous16 [1];
    }
    IAPU.Scanline = 0;
 }
@@ -264,10 +284,21 @@ void S9xSetFilterCoefficient(int tap, int value)
       FilterTapDefinitionBitfield |= 1 << (tap & 7);
 }
 
-void S9xSetSoundADSR(int channel, int attack_rate, int decay_rate,
-                     int sustain_rate, int sustain_level, int release_rate)
+void S9xSetSoundADSR(int channel, int attack_ind, int decay_ind, int sustain_ind, int sustain_level, int release_rate)
 {
+   int attack_rate  = AttackRate [attack_ind];
+   int decay_rate   = DecayRate [decay_ind];
+   int sustain_rate = SustainRate [sustain_ind];
+
+   // Hack for ROMs that use a very short attack rate, key on a
+   // channel, then switch to decay mode. e.g. Final Fantasy II.
+   if(attack_rate == 1)
+      attack_rate = 0;
+
    Channel* ch = &SoundData.channels[channel];
+   ch->env_ind_attack = attack_ind;
+   ch->env_ind_decay = decay_ind;
+   ch->env_ind_sustain = sustain_ind;
    ch->attack_rate = attack_rate;
    ch->decay_rate = decay_rate;
    ch->sustain_rate = sustain_rate;
@@ -277,15 +308,14 @@ void S9xSetSoundADSR(int channel, int attack_rate, int decay_rate,
    switch (SoundData.channels[channel].state)
    {
    case SOUND_ATTACK:
-      S9xSetEnvRate(ch, attack_rate, 1, 127);
+      S9xSetEnvRate(ch, attack_rate, 1, 127, 0);
       break;
-
    case SOUND_DECAY:
       S9xSetEnvRate(ch, decay_rate, -1,
-                    (MAX_ENVELOPE_HEIGHT * (sustain_level + 1)) >> 3);
+                    (MAX_ENVELOPE_HEIGHT * (sustain_level + 1)) >> 3, 1 << 28);
       break;
    case SOUND_SUSTAIN:
-      S9xSetEnvRate(ch, sustain_rate, -1, 0);
+      S9xSetEnvRate(ch, sustain_rate, -1, 0, 2 << 28);
       break;
    }
 }
@@ -320,14 +350,13 @@ int S9xGetEnvelopeHeight(int channel)
    return (0);
 }
 
-void S9xSetSoundFrequency(int channel, int hertz)
+void S9xSetSoundFrequency(int channel, int hertz) // hertz [0~64K<<1]
 {
    if (so.playback_rate)
    {
       if (SoundData.channels[channel].type == SOUND_NOISE)
          hertz = NoiseFreq [APU.DSP [APU_FLG] & 0x1f];
-      SoundData.channels[channel].frequency = (int)
-                                              (((int64_t) hertz * FIXED_POINT) / so.playback_rate);
+      SoundData.channels[channel].frequency = (hertz * so.freqbase) >> 11;
    }
 }
 
@@ -345,13 +374,11 @@ void S9xSetSoundType(int channel, int type_of_sound)
 void DecodeBlock(Channel* ch)
 {
    int32_t out;
-   unsigned char filter;
+   uint8_t filter;
    unsigned char shift;
    signed char sample1, sample2;
-   unsigned char i;
-   bool invalid_header;
 
-   if (ch->block_pointer > 0x10000 - 9)
+   if (ch->block_pointer >= 0x10000 - 9)
    {
       ch->last_block = true;
       ch->loop = false;
@@ -359,274 +386,88 @@ void DecodeBlock(Channel* ch)
       return;
    }
 
-   if (Settings.EightBitConsoleSound)
-   {
-      signed char* compressed = (signed char*) &IAPU.RAM [ch->block_pointer];
-
-      filter = *compressed;
-      if ((ch->last_block = filter & 1))
-         ch->loop = (filter & 2) != 0;
-
-      int16_t interim[16];
-      uint8_t interim_byte = 0;
-
-      compressed++;
+   int8_t* compressed = (int8_t*) &IAPU.RAM [ch->block_pointer];
 
-      // Seperate out the header parts used for decoding
+   filter = *compressed;
+   if ((ch->last_block = filter & 1))
+      ch->loop = (filter & 2) != 0;
 
-      shift = filter >> 4;
+   int16_t* raw = ch->block = ch->decoded;
+   unsigned int i;
 
-      // Header validity check: if range(shift) is over 12, ignore
-      // all bits of the data for that block except for the sign bit of each
-      invalid_header = (shift >= 0xD);
+   compressed++;
 
-      filter = filter & 0x0c;
-
-      int32_t prev0 = ch->previous [0];
-      int32_t prev1 = ch->previous [1];
-
-      int16_t amplitude = 0;
+   int32_t prev0 = ch->previous [0];
+   int32_t prev1 = ch->previous [1];
+   shift = filter >> 4;
 
+   switch ((filter >> 2) & 3)
+   {
+   case 0:
       for (i = 8; i != 0; i--)
       {
          sample1 = *compressed++;
          sample2 = sample1 << 4;
-         //Sample 2 = Bottom Nibble, Sign Extended.
          sample2 >>= 4;
-         //Sample 1 = Top Nibble, shifted down and Sign Extended.
          sample1 >>= 4;
-         if (invalid_header)
-         {
-            sample1 >>= 3;
-            sample2 >>= 3;
-         }
-
-         int nybblesmp;
-         for (nybblesmp = 0; nybblesmp < 2; nybblesmp++)
-         {
-            out = (((nybblesmp) ? sample2 : sample1) << shift);
-            out >>= 1;
-
-            switch (filter)
-            {
-            case 0x00:
-               // Method0 - [Smp]
-               break;
-
-            case 0x04:
-               // Method1 - [Delta]+[Smp-1](15/16)
-               out += (prev0 >> 1) + ((-prev0) >> 5);
-               break;
-
-            case 0x08:
-               // Method2 - [Delta]+[Smp-1](61/32)-[Smp-2](15/16)
-               out += (prev0) + ((-(prev0 + (prev0 >> 1))) >> 5) - (prev1 >> 1) + (prev1 >> 5);
-               break;
-
-            default:
-               // Method3 - [Delta]+[Smp-1](115/64)-[Smp-2](13/16)
-               out += (prev0) + ((-(prev0 + (prev0 << 2) + (prev0 << 3))) >> 7) -
-                      (prev1 >> 1) + ((prev1 + (prev1 >> 1)) >> 4);
-               break;
-
-            }
-            CLIP16(out);
-            int16_t result = (int16_t)(out << 1);
-            if (abs(result) > amplitude)
-               amplitude = abs(result);
-            interim[interim_byte++] = out;
-            prev1 = (int16_t)prev0;
-            prev0 = (int16_t)(out << 1);
-         }
-      }
-      ch->previous [0] = prev0;
-      ch->previous [1] = prev1;
-
-      int32_t total_deviation_from_previous = 0;
-      for (i = 1; i < 16; i++)
-         total_deviation_from_previous += abs(interim[i] - interim[i - 1]);
-      if (total_deviation_from_previous >= (int32_t) amplitude * 4)
-      {
-         /* Looks like noise. Generate noise. */
-         for (i = 0; i < 16; i++)
-         {
-            int feedback = (noise_gen << 13) ^ (noise_gen << 14);
-            noise_gen = (feedback & 0x4000) ^ (noise_gen >> 1);
-            ch->decoded[i] = (noise_gen << 17) >> 17;
-         }
-      }
-      else if (interim[0] < interim[1] && interim[1] < interim[2]
-               && interim[2] < interim[3]
-               && interim[4] > interim[5] && interim[5] > interim[6]
-               && interim[6] > interim[7] && interim[7] > interim[8]
-               && interim[8] > interim[9] && interim[9] > interim[10]
-               && interim[10] > interim[11]
-               && interim[12] < interim[13] && interim[13] < interim[14]
-               && interim[14] < interim[15])
-      {
-         /* Looks like a sine or triangle wave. Make it a
-          * triangle wave with an amplitude equivalent to that
-          * of the highest amplitude sample of the block. */
-         ch->decoded[0] =  ch->decoded[8]  = 0;
-         ch->decoded[1] =  ch->decoded[7]  = amplitude / 4;
-         ch->decoded[2] =  ch->decoded[6]  = amplitude / 2;
-         ch->decoded[3] =  ch->decoded[5]  = amplitude * 3 / 4;
-         ch->decoded[4] =  amplitude;
-         ch->decoded[9] =  ch->decoded[15] = -(amplitude / 4);
-         ch->decoded[10] = ch->decoded[14] = -(amplitude / 2);
-         ch->decoded[11] = ch->decoded[13] = -(amplitude * 3 / 4);
-         ch->decoded[12] = -amplitude;
-      }
-      else if (interim[0] > interim[1] && interim[1] > interim[2]
-               && interim[2] > interim[3]
-               && interim[4] < interim[5] && interim[5] < interim[6]
-               && interim[6] < interim[7] && interim[7] < interim[8]
-               && interim[8] < interim[9] && interim[9] < interim[10]
-               && interim[10] < interim[11]
-               && interim[12] > interim[13] && interim[13] > interim[14]
-               && interim[14] > interim[15])
-      {
-         /* Inverted triangle wave. */
-         ch->decoded[0] =  ch->decoded[8]  = 0;
-         ch->decoded[1] =  ch->decoded[7]  = -(amplitude / 4);
-         ch->decoded[2] =  ch->decoded[6]  = -(amplitude / 2);
-         ch->decoded[3] =  ch->decoded[5]  = -(amplitude * 3 / 4);
-         ch->decoded[4] = -amplitude;
-         ch->decoded[9] =  ch->decoded[15] = amplitude / 4;
-         ch->decoded[10] = ch->decoded[14] = amplitude / 2;
-         ch->decoded[11] = ch->decoded[13] = amplitude * 3 / 4;
-         ch->decoded[12] = amplitude;
-      }
-      else if (interim[0] < interim[1] && interim[1] < interim[2]
-               && interim[2] < interim[3] && interim[3] < interim[4]
-               && interim[4] < interim[5] && interim[5] < interim[6]
-               && interim[6] < interim[7]
-               && interim[8] > interim[9] && interim[9] > interim[10]
-               && interim[10] > interim[11] && interim[11] > interim[12]
-               && interim[12] > interim[13] && interim[13] > interim[14]
-               && interim[14] > interim[15])
-      {
-         /* Looks like a V wave. Make it a half-triangle wave
-          * with an amplitude equivalent to that
-          * of the highest amplitude sample of the block. */
-         ch->decoded[0] =  0;
-         ch->decoded[1] =  ch->decoded[15] = amplitude / 8;
-         ch->decoded[2] =  ch->decoded[14] = amplitude / 4;
-         ch->decoded[3] =  ch->decoded[13] = amplitude * 3 / 8;
-         ch->decoded[4] =  ch->decoded[12] = amplitude / 2;
-         ch->decoded[5] =  ch->decoded[11] = amplitude * 5 / 8;
-         ch->decoded[6] =  ch->decoded[10] = amplitude * 3 / 4;
-         ch->decoded[7] =  ch->decoded[9]  = amplitude * 7 / 8;
-         ch->decoded[8] =  amplitude;
+         *raw++ = ((int32_t) sample1 << shift);
+         *raw++ = ((int32_t) sample2 << shift);
       }
-      else if (interim[0] > interim[1] && interim[1] > interim[2]
-               && interim[2] > interim[3] && interim[3] > interim[4]
-               && interim[4] > interim[5] && interim[5] > interim[6]
-               && interim[6] > interim[7]
-               && interim[8] < interim[9] && interim[9] < interim[10]
-               && interim[10] < interim[11] && interim[11] < interim[12]
-               && interim[12] < interim[13] && interim[13] < interim[14]
-               && interim[14] < interim[15])
+      prev1 = *(raw - 2);
+      prev0 = *(raw - 1);
+      break;
+   case 1:
+      for (i = 8; i != 0; i--)
       {
-         /* Inverted V wave. */
-         ch->decoded[0] =  0;
-         ch->decoded[1] =  ch->decoded[15] = -(amplitude / 8);
-         ch->decoded[2] =  ch->decoded[14] = -(amplitude / 4);
-         ch->decoded[3] =  ch->decoded[13] = -(amplitude * 3 / 8);
-         ch->decoded[4] =  ch->decoded[12] = -(amplitude / 2);
-         ch->decoded[5] =  ch->decoded[11] = -(amplitude * 5 / 8);
-         ch->decoded[6] =  ch->decoded[10] = -(amplitude * 3 / 4);
-         ch->decoded[7] =  ch->decoded[9]  = -(amplitude * 7 / 8);
-         ch->decoded[8] =  -amplitude;
+         sample1 = *compressed++;
+         sample2 = sample1 << 4;
+         sample2 >>= 4;
+         sample1 >>= 4;
+         prev0 = (int16_t) prev0;
+         *raw++ = prev1 = ((int32_t) sample1 << shift) + prev0 - (prev0 >> 4);
+         prev1 = (int16_t) prev1;
+         *raw++ = prev0 = ((int32_t) sample2 << shift) + prev1 - (prev1 >> 4);
       }
-      else
+      break;
+   case 2:
+      for (i = 8; i != 0; i--)
       {
-         // Make it a square wave with an amplitude equivalent to that
-         // of the highest amplitude sample of the block.
-         // But actually put half of the amplitude, because
-         // square waves are just loud.
-         for (i = 0; i < 8; i++)
-            ch->decoded[i] = amplitude / 2;
-         for (i = 8; i < 16; i++)
-            ch->decoded[i] = -(amplitude / 2);
+         sample1 = *compressed++;
+         sample2 = sample1 << 4;
+         sample2 >>= 4;
+         sample1 >>= 4;
+         out = (sample1 << shift) - prev1 + (prev1 >> 4);
+         prev1 = (int16_t) prev0;
+         prev0 &= ~3;
+         *raw++ = prev0 = out + (prev0 << 1) - (prev0 >> 5) - (prev0 >> 4);
+         out = (sample2 << shift) - prev1 + (prev1 >> 4);
+         prev1 = (int16_t) prev0;
+         prev0 &= ~3;
+         *raw++ = prev0 = out + (prev0 << 1) - (prev0 >> 5) - (prev0 >> 4);
       }
-   }
-   else
-   {
-      signed char* compressed = (signed char*) &IAPU.RAM [ch->block_pointer];
-
-      filter = *compressed;
-      if ((ch->last_block = filter & 1))
-         ch->loop = (filter & 2) != 0;
-
-      compressed++;
-      int16_t* raw = ch->block = ch->decoded;
-
-      // Seperate out the header parts used for decoding
-
-      shift = filter >> 4;
-
-      // Header validity check: if range(shift) is over 12, ignore
-      // all bits of the data for that block except for the sign bit of each
-      invalid_header = (shift >= 0xD);
-
-      filter = filter & 0x0c;
-
-      int32_t prev0 = ch->previous [0];
-      int32_t prev1 = ch->previous [1];
-
+      break;
+   case 3:
       for (i = 8; i != 0; i--)
       {
          sample1 = *compressed++;
          sample2 = sample1 << 4;
-         //Sample 2 = Bottom Nibble, Sign Extended.
          sample2 >>= 4;
-         //Sample 1 = Top Nibble, shifted down and Sign Extended.
          sample1 >>= 4;
-         if (invalid_header)
-         {
-            sample1 >>= 3;
-            sample2 >>= 3;
-         }
-
-         int nybblesmp;
-         for (nybblesmp = 0; nybblesmp < 2; nybblesmp++)
-         {
-            out = (((nybblesmp) ? sample2 : sample1) << shift);
-            out >>= 1;
-
-            switch (filter)
-            {
-            case 0x00:
-               // Method0 - [Smp]
-               break;
-
-            case 0x04:
-               // Method1 - [Delta]+[Smp-1](15/16)
-               out += (prev0 >> 1) + ((-prev0) >> 5);
-               break;
-
-            case 0x08:
-               // Method2 - [Delta]+[Smp-1](61/32)-[Smp-2](15/16)
-               out += (prev0) + ((-(prev0 + (prev0 >> 1))) >> 5) - (prev1 >> 1) + (prev1 >> 5);
-               break;
-
-            default:
-               // Method3 - [Delta]+[Smp-1](115/64)-[Smp-2](13/16)
-               out += (prev0) + ((-(prev0 + (prev0 << 2) + (prev0 << 3))) >> 7) -
-                      (prev1 >> 1) + ((prev1 + (prev1 >> 1)) >> 4);
-               break;
-
-            }
-            CLIP16(out);
-            *raw++ = (int16_t)(out << 1);
-            prev1 = (int16_t)prev0;
-            prev0 = (int16_t)(out << 1);
-         }
+         out = (sample1 << shift);
+         out = out - prev1 + (prev1 >> 3) + (prev1 >> 4);
+         prev1 = (int16_t) prev0;
+         prev0 &= ~3;
+         *raw++ = prev0 = out + (prev0 << 1) - (prev0 >> 3) - (prev0 >> 4) - (prev1 >> 6);
+         out = (sample2 << shift);
+         out = out - prev1 + (prev1 >> 3) + (prev1 >> 4);
+         prev1 = (int16_t) prev0;
+         prev0 &= ~3;
+         *raw++ = prev0 = out + (prev0 << 1) - (prev0 >> 3) - (prev0 >> 4) - (prev1 >> 6);
       }
-      ch->previous [0] = prev0;
-      ch->previous [1] = prev1;
+      break;
    }
+   ch->previous [0] = prev0;
+   ch->previous [1] = prev1;
    ch->block_pointer += 9;
 }
 
@@ -641,7 +482,7 @@ static inline void MixStereo(int sample_count)
    {
       Channel* ch = &SoundData.channels[J];
 
-      if (ch->state == SOUND_SILENT || !(so.sound_switch & (1 << J)))
+      if (ch->state == SOUND_SILENT)
          continue;
 
       int32_t VL, VR;
@@ -698,15 +539,13 @@ static inline void MixStereo(int sample_count)
                   if (ch->sustain_level != 8)
                   {
                      S9xSetEnvRate(ch, ch->decay_rate, -1,
-                                   (MAX_ENVELOPE_HEIGHT * ch->sustain_level)
-                                   >> 3);
+                                   (MAX_ENVELOPE_HEIGHT * ch->sustain_level) >> 3, 1 << 28);
                      break;
                   }
                   ch->state = SOUND_SUSTAIN;
-                  S9xSetEnvRate(ch, ch->sustain_rate, -1, 0);
+                  S9xSetEnvRate(ch, ch->sustain_rate, -1, 0, 2 << 28);
                }
                break;
-
             case SOUND_DECAY:
                while (ch->env_error >= FIXED_POINT)
                {
@@ -722,10 +561,9 @@ static inline void MixStereo(int sample_count)
                      goto stereo_exit;
                   }
                   ch->state = SOUND_SUSTAIN;
-                  S9xSetEnvRate(ch, ch->sustain_rate, -1, 0);
+                  S9xSetEnvRate(ch, ch->sustain_rate, -1, 0, 2 << 28);
                }
                break;
-
             case SOUND_SUSTAIN:
                while (ch->env_error >= FIXED_POINT)
                {
@@ -739,7 +577,6 @@ static inline void MixStereo(int sample_count)
                   goto stereo_exit;
                }
                break;
-
             case SOUND_RELEASE:
                while (ch->env_error >= FIXED_POINT)
                {
@@ -753,7 +590,6 @@ static inline void MixStereo(int sample_count)
                   goto stereo_exit;
                }
                break;
-
             case SOUND_INCREASE_LINEAR:
                ch->env_error &= FIXED_POINT_REMAINDER;
                ch->envx += step << 1;
@@ -765,10 +601,9 @@ static inline void MixStereo(int sample_count)
                   ch->envxx = 127 << ENVX_SHIFT;
                   ch->state = SOUND_GAIN;
                   ch->mode = MODE_GAIN;
-                  S9xSetEnvRate(ch, 0, -1, 0);
+                  S9xSetEnvRate(ch, 0, -1, 0, 0);
                }
                break;
-
             case SOUND_INCREASE_BENT_LINE:
                if (ch->envx >= (MAX_ENVELOPE_HEIGHT * 3) / 4)
                {
@@ -792,10 +627,9 @@ static inline void MixStereo(int sample_count)
                   ch->envxx = 127 << ENVX_SHIFT;
                   ch->state = SOUND_GAIN;
                   ch->mode = MODE_GAIN;
-                  S9xSetEnvRate(ch, 0, -1, 0);
+                  S9xSetEnvRate(ch, 0, -1, 0, 0);
                }
                break;
-
             case SOUND_DECREASE_LINEAR:
                ch->env_error &= FIXED_POINT_REMAINDER;
                ch->envx -= step << 1;
@@ -806,7 +640,6 @@ static inline void MixStereo(int sample_count)
                   goto stereo_exit;
                }
                break;
-
             case SOUND_DECREASE_EXPONENTIAL:
                while (ch->env_error >= FIXED_POINT)
                {
@@ -820,9 +653,8 @@ static inline void MixStereo(int sample_count)
                   goto stereo_exit;
                }
                break;
-
             case SOUND_GAIN:
-               S9xSetEnvRate(ch, 0, -1, 0);
+               S9xSetEnvRate(ch, 0, -1, 0, 0);
                break;
             }
             ch-> left_vol_level = (ch->envx * ch->volume_left) / 128;
@@ -916,6 +748,10 @@ static inline void MixStereo(int sample_count)
 
          MixBuffer [I    ] += VL;
          MixBuffer [I + 1] += VR;
+
+         if (!ch->echo_buf_ptr)
+            continue;
+         
          ch->echo_buf_ptr [I    ] += VL;
          ch->echo_buf_ptr [I + 1] += VR;
       }
@@ -924,11 +760,17 @@ stereo_exit:
    }
 }
 
-void S9xMixSamples(uint8_t* buffer, int sample_count)
+void S9xMixSamples(uint16_t* buffer, int sample_count)
 {
    int J;
    int I;
 
+   if (so.mute_sound)
+   {
+      memset(buffer, 0, sample_count * sizeof(buffer [0]));
+      return;
+   }
+
    if (SoundData.echo_enable)
       memset(EchoBuffer, 0, sample_count * sizeof(EchoBuffer [0]));
    memset(MixBuffer, 0, sample_count * sizeof(MixBuffer [0]));
@@ -944,19 +786,14 @@ void S9xMixSamples(uint8_t* buffer, int sample_count)
          for (J = 0; J < sample_count; J++)
          {
             int E = Echo [SoundData.echo_ptr];
+            Echo[SoundData.echo_ptr++] = (E * SoundData.echo_feedback) / 128 + EchoBuffer [J];
 
-            Echo [SoundData.echo_ptr] = (E * SoundData.echo_feedback) / 128 +
-                                        EchoBuffer [J];
-
-            if ((SoundData.echo_ptr += 1) >= SoundData.echo_buffer_size)
+            if (SoundData.echo_ptr >= SoundData.echo_buffer_size)
                SoundData.echo_ptr = 0;
 
-            I = (MixBuffer [J] *
-                 SoundData.master_volume [J & 1] +
-                 E * SoundData.echo_volume [J & 1]) / VOL_DIV16;
-
+            I = (MixBuffer[J] * SoundData.master_volume [J & 1] + E * SoundData.echo_volume [J & 1]) / VOL_DIV16;
             CLIP16(I);
-            ((int16_t*) buffer)[J] = I;
+            buffer[J] = I;
          }
       }
       else
@@ -964,39 +801,26 @@ void S9xMixSamples(uint8_t* buffer, int sample_count)
          // ... with filter defined.
          for (J = 0; J < sample_count; J++)
          {
-            int E = Echo [SoundData.echo_ptr];
-
-            Loop [(Z - 0) & 15] = E;
-            E =  E                    * FilterTaps [0];
-            if (FilterTapDefinitionBitfield & 0x02) E += Loop [(Z -  2) & 15] *
-                     FilterTaps [1];
-            if (FilterTapDefinitionBitfield & 0x04) E += Loop [(Z -  4) & 15] *
-                     FilterTaps [2];
-            if (FilterTapDefinitionBitfield & 0x08) E += Loop [(Z -  6) & 15] *
-                     FilterTaps [3];
-            if (FilterTapDefinitionBitfield & 0x10) E += Loop [(Z -  8) & 15] *
-                     FilterTaps [4];
-            if (FilterTapDefinitionBitfield & 0x20) E += Loop [(Z - 10) & 15] *
-                     FilterTaps [5];
-            if (FilterTapDefinitionBitfield & 0x40) E += Loop [(Z - 12) & 15] *
-                     FilterTaps [6];
-            if (FilterTapDefinitionBitfield & 0x80) E += Loop [(Z - 14) & 15] *
-                     FilterTaps [7];
+            Loop [(Z - 0) & 15] = Echo [SoundData.echo_ptr];
+            int                                     E =  Loop [(Z -  0) & 15] * FilterTaps [0];
+            if (FilterTapDefinitionBitfield & 0x02) E += Loop [(Z -  2) & 15] * FilterTaps [1];
+            if (FilterTapDefinitionBitfield & 0x04) E += Loop [(Z -  4) & 15] * FilterTaps [2];
+            if (FilterTapDefinitionBitfield & 0x08) E += Loop [(Z -  6) & 15] * FilterTaps [3];
+            if (FilterTapDefinitionBitfield & 0x10) E += Loop [(Z -  8) & 15] * FilterTaps [4];
+            if (FilterTapDefinitionBitfield & 0x20) E += Loop [(Z - 10) & 15] * FilterTaps [5];
+            if (FilterTapDefinitionBitfield & 0x40) E += Loop [(Z - 12) & 15] * FilterTaps [6];
+            if (FilterTapDefinitionBitfield & 0x80) E += Loop [(Z - 14) & 15] * FilterTaps [7];
             E /= 128;
             Z++;
 
-            Echo [SoundData.echo_ptr] = (E * SoundData.echo_feedback) / 128 +
-                                        EchoBuffer [J];
+            Echo[SoundData.echo_ptr++] = (E * SoundData.echo_feedback) / 128 + EchoBuffer[J];
 
-            if ((SoundData.echo_ptr += 1) >= SoundData.echo_buffer_size)
+            if (SoundData.echo_ptr >= SoundData.echo_buffer_size)
                SoundData.echo_ptr = 0;
 
-            I = (MixBuffer [J] *
-                 SoundData.master_volume [J & 1] +
-                 E * SoundData.echo_volume [J & 1]) / VOL_DIV16;
-
+            I = (MixBuffer[J] * SoundData.master_volume [J & 1] + E * SoundData.echo_volume [J & 1]) / VOL_DIV16;
             CLIP16(I);
-            ((int16_t*) buffer)[J] = I;
+            buffer[J] = I;
          }
       }
    }
@@ -1005,14 +829,11 @@ void S9xMixSamples(uint8_t* buffer, int sample_count)
       // 16-bit mono or stereo sound, no echo
       for (J = 0; J < sample_count; J++)
       {
-         I = (MixBuffer [J] *
-              SoundData.master_volume [J & 1]) / VOL_DIV16;
-
+         I = (MixBuffer[J] * SoundData.master_volume [J & 1]) / VOL_DIV16;
          CLIP16(I);
-         ((int16_t*) buffer)[J] = I;
+         buffer[J] = I;
       }
    }
-
 }
 
 void S9xResetSound(bool full)
@@ -1041,6 +862,10 @@ void S9xResetSound(bool full)
       SoundData.channels[i].sustain_rate = 0;
       SoundData.channels[i].release_rate = 0;
       SoundData.channels[i].sustain_level = 0;
+      // notaz
+      SoundData.channels[i].env_ind_attack = 0;
+      SoundData.channels[i].env_ind_decay = 0;
+      SoundData.channels[i].env_ind_sustain = 0;
       SoundData.echo_ptr = 0;
       SoundData.echo_feedback = 0;
       SoundData.echo_buffer_size = 1;
@@ -1055,10 +880,6 @@ void S9xResetSound(bool full)
    FilterTaps [7] = 0;
    FilterTapDefinitionBitfield = 0;
    noise_gen = 1;
-   so.sound_switch = 255;
-   so.samples_mixed_so_far = 0;
-   so.play_position = 0;
-   so.err_counter = 0;
 
    if (full)
    {
@@ -1066,9 +887,6 @@ void S9xResetSound(bool full)
       SoundData.echo_write_enabled = 0;
       SoundData.echo_channel_enable = 0;
       SoundData.pitch_mod = 0;
-      SoundData.dummy[0] = 0;
-      SoundData.dummy[1] = 0;
-      SoundData.dummy[2] = 0;
       SoundData.master_volume[0] = 0;
       SoundData.master_volume[1] = 0;
       SoundData.echo_volume[0] = 0;
@@ -1077,18 +895,46 @@ void S9xResetSound(bool full)
    }
 
    SoundData.master_volume [0] = SoundData.master_volume [1] = 127;
-   if (so.playback_rate)
-      so.err_rate = (uint32_t)(FIXED_POINT * SNES_SCANLINE_TIME /
-                             (1.0 / so.playback_rate));
-   else
-      so.err_rate = 0;
+   so.mute_sound = true;
 }
 
 void S9xSetPlaybackRate(uint32_t playback_rate)
 {
    so.playback_rate = playback_rate;
-   so.err_rate = (uint32_t)(SNES_SCANLINE_TIME * FIXED_POINT / (1.0 /
-                          (double) so.playback_rate));
+
+   if (playback_rate)
+   {
+      // notaz: calculate a value (let's call it freqbase) to simplify channel freq calculations later.
+      so.freqbase = (FIXED_POINT << 11) / (playback_rate * 33 / 32);
+      // now precalculate env rates for S9xSetEnvRate
+      static int steps [] =
+      {
+         0, 64, 619, 619, 128, 1, 64, 55, 64, 619
+      };
+
+      int i, u;
+
+      for (u = 0 ; u < 10 ; u++)
+      {
+         int64_t fp1000su = ((int64_t) FIXED_POINT * 1000 * steps[u]);
+
+         for (i = 0 ; i < 16 ; i++)
+            AttackERate[i][u] = (uint32_t) (fp1000su / (AttackRate[i] * playback_rate));
+
+         for (i = 0 ; i < 8 ; i++)
+            DecayERate[i][u]  = (uint32_t) (fp1000su / (DecayRate[i]  * playback_rate));
+
+         for (i = 0 ; i < 32 ; i++)
+         {
+            SustainERate[i][u] = (uint32_t) (fp1000su / (SustainRate[i] * playback_rate));
+            IncreaseERate[i][u] = (uint32_t) (fp1000su / (IncreaseRate[i] * playback_rate));
+            DecreaseERateExp[i][u] = (uint32_t) (fp1000su / (DecreaseRateExp[i] / 2 * playback_rate));
+         }
+
+         KeyOffERate[u] = (uint32_t) (fp1000su / (8 * playback_rate));
+      }
+   }
+
    S9xSetEchoDelay(APU.DSP [APU_EDL] & 0xf);
    int i;
    for (i = 0; i < 8; i++)
@@ -1097,13 +943,8 @@ void S9xSetPlaybackRate(uint32_t playback_rate)
 
 bool S9xInitSound()
 {
-   so.sound_fd = -1;
-   so.sound_switch = 255;
-
    so.playback_rate = 0;
-   so.buffer_size = 0;
-   so.encoded = false;
-
+   S9xResetSound(true);
    return (1);
 }
 
@@ -1120,20 +961,9 @@ bool S9xSetSoundMode(int channel, int mode)
          return (true);
       }
       break;
-
    case MODE_DECREASE_LINEAR:
    case MODE_DECREASE_EXPONENTIAL:
    case MODE_GAIN:
-      if (ch->mode != MODE_RELEASE)
-      {
-         ch->mode = mode;
-         if (ch->state != SOUND_SILENT)
-            ch->state = mode;
-
-         return (true);
-      }
-      break;
-
    case MODE_INCREASE_LINEAR:
    case MODE_INCREASE_BENT_LINE:
       if (ch->mode != MODE_RELEASE)
@@ -1141,11 +971,9 @@ bool S9xSetSoundMode(int channel, int mode)
          ch->mode = mode;
          if (ch->state != SOUND_SILENT)
             ch->state = mode;
-
          return (true);
       }
       break;
-
    case MODE_ADSR:
       if (ch->mode == MODE_NONE || ch->mode == MODE_ADSR)
       {
@@ -1157,11 +985,6 @@ bool S9xSetSoundMode(int channel, int mode)
    return (false);
 }
 
-void S9xSetSoundControl(int sound_switch)
-{
-   so.sound_switch = sound_switch;
-}
-
 void S9xPlaySample(int channel)
 {
    Channel* ch = &SoundData.channels[channel];
@@ -1202,14 +1025,14 @@ void S9xPlaySample(int channel)
          {
             ch->state = SOUND_SUSTAIN;
             ch->envx = (MAX_ENVELOPE_HEIGHT * ch->sustain_level) >> 3;
-            S9xSetEnvRate(ch, ch->sustain_rate, -1, 0);
+            S9xSetEnvRate(ch, ch->sustain_rate, -1, 0, 2 << 28);
          }
          else
          {
             ch->state = SOUND_DECAY;
             ch->envx = MAX_ENVELOPE_HEIGHT;
             S9xSetEnvRate(ch, ch->decay_rate, -1,
-                          (MAX_ENVELOPE_HEIGHT * ch->sustain_level) >> 3);
+                          (MAX_ENVELOPE_HEIGHT * ch->sustain_level) >> 3, 1 << 28);
          }
          ch-> left_vol_level = (ch->envx * ch->volume_left) / 128;
          ch->right_vol_level = (ch->envx * ch->volume_right) / 128;
@@ -1220,31 +1043,25 @@ void S9xPlaySample(int channel)
          ch->envx = 0;
          ch->left_vol_level = 0;
          ch->right_vol_level = 0;
-         S9xSetEnvRate(ch, ch->attack_rate, 1, MAX_ENVELOPE_HEIGHT);
+         S9xSetEnvRate(ch, ch->attack_rate, 1, MAX_ENVELOPE_HEIGHT, 0);
       }
       ch->envxx = ch->envx << ENVX_SHIFT;
       break;
-
    case MODE_GAIN:
       ch->state = SOUND_GAIN;
       break;
-
    case MODE_INCREASE_LINEAR:
       ch->state = SOUND_INCREASE_LINEAR;
       break;
-
    case MODE_INCREASE_BENT_LINE:
       ch->state = SOUND_INCREASE_BENT_LINE;
       break;
-
    case MODE_DECREASE_LINEAR:
       ch->state = SOUND_DECREASE_LINEAR;
       break;
-
    case MODE_DECREASE_EXPONENTIAL:
       ch->state = SOUND_DECREASE_EXPONENTIAL;
       break;
-
    default:
       break;
    }
diff --git a/source/soundux.h b/source/soundux.h
index 411e334..e4cf3ad 100644
--- a/source/soundux.h
+++ b/source/soundux.h
@@ -24,23 +24,17 @@ enum { MODE_NONE = SOUND_SILENT, MODE_ADSR, MODE_RELEASE = SOUND_RELEASE,
 #define SOUND_DECODE_LENGTH 16
 
 #define NUM_CHANNELS    8
-#define SOUND_BUFFER_SIZE (1024 * 16)
+#define SOUND_BUFFER_SIZE (2 * 44100 / 50)
 #define MAX_BUFFER_SIZE SOUND_BUFFER_SIZE
-#define SOUND_BUFFER_SIZE_MASK (SOUND_BUFFER_SIZE - 1)
 
 #define SOUND_BUFS      4
 
 typedef struct
 {
-   int      sound_fd;
-   int      sound_switch;
    int      playback_rate;
    int      buffer_size;
-   bool     encoded;
-   int32_t  samples_mixed_so_far;
-   int32_t  play_position;
-   uint32_t err_counter;
-   uint32_t err_rate;
+   uint32_t freqbase; // notaz
+   bool     mute_sound;
 } SoundStatus;
 
 SoundStatus so;
@@ -82,8 +76,12 @@ typedef struct
    int16_t  next_sample;
    int32_t  interpolate;
    int32_t  previous [2];
+   // notaz
+   uint8_t  env_ind_attack;
+   uint8_t  env_ind_decay;
+   uint8_t  env_ind_sustain;
    // Just incase they are needed in the future, for snapshot compatibility.
-   uint32_t dummy [8];
+   uint8_t  dummy [29];
 } Channel;
 
 typedef struct
@@ -105,18 +103,14 @@ typedef struct
 
 SSoundData SoundData;
 
-void S9xSetEightBitConsoleSound(bool Enabled);
-
 void S9xSetSoundVolume(int channel, int16_t volume_left, int16_t volume_right);
 void S9xSetSoundFrequency(int channel, int hertz);
 void S9xSetSoundHertz(int channel, int hertz);
 void S9xSetSoundType(int channel, int type_of_sound);
 void S9xSetMasterVolume(int16_t master_volume_left, int16_t master_volume_right);
 void S9xSetEchoVolume(int16_t echo_volume_left, int16_t echo_volume_right);
-void S9xSetSoundControl(int sound_switch);
 void S9xSetEnvelopeHeight(int channel, int height);
-void S9xSetSoundADSR(int channel, int attack, int decay, int sustain,
-                     int sustain_level, int release);
+void S9xSetSoundADSR(int channel, int attack, int decay, int sustain, int sustain_level, int release);
 void S9xSetSoundKeyOff(int channel);
 void S9xSetSoundDecayMode(int channel);
 void S9xSetSoundAttachMode(int channel);
@@ -127,8 +121,7 @@ void S9xSetEchoDelay(int byte);
 void S9xSetEchoWriteEnable(uint8_t byte);
 void S9xSetFilterCoefficient(int tap, int value);
 void S9xSetFrequencyModulationEnable(uint8_t byte);
-void S9xSetEnvelopeRate(int channel, uint32_t rate, int direction,
-                        int target);
+void S9xSetEnvelopeRate(int channel, uint32_t rate, int direction, int target, unsigned int mode);
 bool S9xSetSoundMode(int channel, int mode);
 void S9xResetSound(bool full);
 void S9xFixSoundAfterSnapshotLoad();
@@ -137,7 +130,7 @@ void S9xPlaySample(int channel);
 void S9xFixEnvelope(int channel, uint8_t gain, uint8_t adsr1, uint8_t adsr2);
 void S9xStartSample(int channel);
 
-void S9xMixSamples(uint8_t* buffer, int sample_count);
+void S9xMixSamples(uint16_t* buffer, int sample_count);
 bool S9xOpenSoundDevice(int, bool, int);
 void S9xSetPlaybackRate(uint32_t rate);
 #endif
diff --git a/source/spc700.c b/source/spc700.c
index d152bd1..5e50cf3 100644
--- a/source/spc700.c
+++ b/source/spc700.c
@@ -23,6 +23,15 @@ void S9xAPUSetByte(uint8_t, uint32_t address);
 #include "apumem.h"
 #endif
 
+int8_t   Int8 = 0;
+int16_t  Int16 = 0;
+int32_t  Int32 = 0;
+uint8_t  W1;
+uint8_t  W2;
+uint8_t  Work8 = 0;
+uint16_t Work16 = 0;
+uint32_t Work32 = 0;
+
 #define OP1 (*(IAPU.PC + 1))
 #define OP2 (*(IAPU.PC + 2))
 
@@ -56,13 +65,12 @@ void S9xAPUSetByte(uint8_t, uint32_t address);
 #define TCALL(n)\
 {\
     SPC700_PushW (IAPU.PC - IAPU.RAM + 1); \
-    IAPU.PC = IAPU.RAM + (APU.ExtraRAM [((15 - n) << 1)] + \
-        (APU.ExtraRAM [((15 - n) << 1) + 1] << 8)); \
+    IAPU.PC = IAPU.RAM + S9xAPUGetByte(0xffc0 + ((15 - n) << 1)) + \
+        (S9xAPUGetByte(0xffc1 + ((15 - n) << 1)) << 8); \
 }
 
-// XXX: HalfCarry - BJ fixed?
 #define SBC(a,b)\
-int16_t Int16 = (int16_t) (a) - (int16_t) (b) + (int16_t) (APUCheckCarry ()) - 1;\
+Int16 = (int16_t) (a) - (int16_t) (b) + (int16_t) (APUCheckCarry ()) - 1;\
 IAPU._Carry = Int16 >= 0;\
 if ((((a) ^ (b)) & 0x80) && (((a) ^ (uint8_t) Int16) & 0x80))\
     APUSetOverflow ();\
@@ -74,10 +82,8 @@ if(((a) ^ (b) ^ (uint8_t) Int16) & 0x10)\
 (a) = (uint8_t) Int16;\
 APUSetZN8 ((uint8_t) Int16);
 
-// XXX: HalfCarry - BJ fixed?
-// XXX: HalfCarry used Int16 before; trying to fix it with Work16 [Neb]
 #define ADC(a,b)\
-uint16_t Work16 = (a) + (b) + APUCheckCarry();\
+Work16 = (a) + (b) + APUCheckCarry();\
 IAPU._Carry = Work16 >= 0x100; \
 if (~((a) ^ (b)) & ((b) ^ (uint8_t) Work16) & 0x80)\
     APUSetOverflow ();\
@@ -90,7 +96,7 @@ if(((a) ^ (b) ^ (uint8_t) Work16) & 0x10)\
 APUSetZN8 ((uint8_t) Work16);
 
 #define CMP(a,b)\
-int16_t Int16 = (int16_t) (a) - (int16_t) (b);\
+Int16 = (int16_t) (a) - (int16_t) (b);\
 IAPU._Carry = Int16 >= 0;\
 APUSetZN8 ((uint8_t) Int16);
 
@@ -103,12 +109,12 @@ APUSetZN8 ((uint8_t) Int16);
     (b) >>= 1;\
     APUSetZN8 (b);
 #define ROL(b)\
-    uint16_t Work16 = ((b) << 1) | APUCheckCarry (); \
+    Work16 = ((b) << 1) | APUCheckCarry (); \
     IAPU._Carry = Work16 >= 0x100; \
     (b) = (uint8_t) Work16; \
     APUSetZN8 (b);
 #define ROR(b)\
-    uint16_t Work16 = (b) | ((uint16_t) APUCheckCarry () << 8); \
+    Work16 = (b) | ((uint16_t) APUCheckCarry () << 8); \
     IAPU._Carry = (uint8_t) Work16 & 1; \
     Work16 >>= 1; \
     (b) = (uint8_t) Work16; \
@@ -124,11 +130,19 @@ APUSetZN8 ((uint8_t) Int16);
 
 #ifdef FAST_LSB_WORD_ACCESS
 #define SPC700_PushW(w)\
-    *(uint16_t *) (IAPU.RAM + 0xff + IAPU.Registers.S) = w;\
+    if (IAPU.Registers.S == 0) {\
+        *(IAPU.RAM + 0x1ff) = (w);\
+        *(IAPU.RAM + 0x100) = ((w) >> 8);\
+    } else { \
+        *(uint16_t *) (IAPU.RAM + 0xff + IAPU.Registers.S) = w;\
+    }\
     IAPU.Registers.S -= 2;
 #define PopW(w)\
     IAPU.Registers.S += 2;\
-    w = *(uint16_t *) (IAPU.RAM + 0xff + IAPU.Registers.S);
+    if (IAPU.Registers.S == 0)\
+       (w) = *(IAPU.RAM + 0x1ff) | (*(IAPU.RAM + 0x100) << 8);\
+    else \
+       (w) = *(uint16_t *) (IAPU.RAM + 0xff + IAPU.Registers.S);
 #else
 #define SPC700_PushW(w)\
     *(IAPU.RAM + 0xff + IAPU.Registers.S) = w;\
@@ -136,16 +150,19 @@ APUSetZN8 ((uint8_t) Int16);
     IAPU.Registers.S -= 2;
 #define PopW(w)\
     IAPU.Registers.S += 2; \
-    (w) = *(IAPU.RAM + 0xff + IAPU.Registers.S) + (*(IAPU.RAM + 0x100 + IAPU.Registers.S) << 8);
+    if(IAPU.Registers.S == 0) \
+       (w) = *(IAPU.RAM + 0x1ff) | (*(IAPU.RAM + 0x100) << 8); \
+    else \
+       (w) = *(IAPU.RAM + 0xff + IAPU.Registers.S) + (*(IAPU.RAM + 0x100 + IAPU.Registers.S) << 8);
 #endif
 
 #define Relative()\
-    int8_t Int8 = OP1;\
-    int16_t Int16 = (intptr_t) (IAPU.PC + 2 - IAPU.RAM) + Int8;
+    Int8 = OP1;\
+    Int16 = (intptr_t) (IAPU.PC + 2 - IAPU.RAM) + Int8;
 
 #define Relative2()\
-    int8_t Int8 = OP2;\
-    int16_t Int16 = (intptr_t) (IAPU.PC + 3 - IAPU.RAM) + Int8;
+    Int8 = OP2;\
+    Int16 = (intptr_t) (IAPU.PC + 3 - IAPU.RAM) + Int8;
 
 #ifdef FAST_LSB_WORD_ACCESS
 #define IndexedXIndirect()\
@@ -287,7 +304,7 @@ void Apu3F()  // CALL absolute
 
 void Apu4F()  // PCALL $XX
 {
-   uint8_t Work8 = OP1;
+   Work8 = OP1;
    SPC700_PushW(IAPU.PC + 2 - IAPU.RAM);
    IAPU.PC = IAPU.RAM + 0xff00 + Work8;
 }
@@ -381,7 +398,7 @@ void ApuF2()
 }
 
 #define BBS(b) \
-uint8_t Work8 = OP1; \
+Work8 = OP1; \
 Relative2 (); \
 if (S9xAPUGetByteZ (Work8) & (1 << (b))) \
 { \
@@ -432,7 +449,7 @@ void ApuE3()
 }
 
 #define BBC(b) \
-uint8_t Work8 = OP1; \
+Work8 = OP1; \
 Relative2 (); \
 if (!(S9xAPUGetByteZ (Work8) & (1 << (b)))) \
 { \
@@ -527,7 +544,7 @@ void Apu08()
 void Apu09()
 {
    // OR dp(dest),dp(src)
-   uint8_t Work8 = S9xAPUGetByteZ(OP1);
+   Work8 = S9xAPUGetByteZ(OP1);
    Work8 |= S9xAPUGetByteZ(OP2);
    S9xAPUSetByteZ(Work8, OP2);
    APUSetZN8(Work8);
@@ -572,7 +589,7 @@ void Apu17()
 void Apu18()
 {
    // OR dp,#00
-   uint8_t Work8 = OP1;
+   Work8 = OP1;
    Work8 |= S9xAPUGetByteZ(OP2);
    S9xAPUSetByteZ(Work8, OP2);
    APUSetZN8(Work8);
@@ -582,8 +599,7 @@ void Apu18()
 void Apu19()
 {
    // OR (X),(Y)
-   uint8_t Work8 = S9xAPUGetByteZ(IAPU.Registers.X) | S9xAPUGetByteZ(
-                    IAPU.Registers.YA.B.Y);
+   Work8 = S9xAPUGetByteZ(IAPU.Registers.X) | S9xAPUGetByteZ(IAPU.Registers.YA.B.Y);
    APUSetZN8(Work8);
    S9xAPUSetByteZ(Work8, IAPU.Registers.X);
    IAPU.PC++;
@@ -641,14 +657,11 @@ void Apu8A()
 {
    // EOR1 C, membit
    MemBit();
-   if (APUCheckCarry())
+   if (S9xAPUGetByte(IAPU.Address) & (1 << IAPU.Bit))
    {
-      if (S9xAPUGetByte(IAPU.Address) & (1 << IAPU.Bit))
+      if (APUCheckCarry())
          APUClearCarry();
-   }
-   else
-   {
-      if (S9xAPUGetByte(IAPU.Address) & (1 << IAPU.Bit))
+      else
          APUSetCarry();
    }
    IAPU.PC += 3;
@@ -687,7 +700,7 @@ void ApuEA()
 void Apu0B()
 {
    // ASL dp
-   uint8_t Work8 = S9xAPUGetByteZ(OP1);
+   Work8 = S9xAPUGetByteZ(OP1);
    ASL(Work8);
    S9xAPUSetByteZ(Work8, OP1);
    IAPU.PC += 2;
@@ -697,7 +710,7 @@ void Apu0C()
 {
    // ASL abs
    Absolute();
-   uint8_t Work8 = S9xAPUGetByte(IAPU.Address);
+   Work8 = S9xAPUGetByte(IAPU.Address);
    ASL(Work8);
    S9xAPUSetByte(Work8, IAPU.Address);
    IAPU.PC += 3;
@@ -706,7 +719,7 @@ void Apu0C()
 void Apu1B()
 {
    // ASL dp+X
-   uint8_t Work8 = S9xAPUGetByteZ(OP1 + IAPU.Registers.X);
+   Work8 = S9xAPUGetByteZ(OP1 + IAPU.Registers.X);
    ASL(Work8);
    S9xAPUSetByteZ(Work8, OP1 + IAPU.Registers.X);
    IAPU.PC += 2;
@@ -785,9 +798,9 @@ void Apu0E()
 {
    // TSET1 abs
    Absolute();
-   uint8_t Work8 = S9xAPUGetByte(IAPU.Address);
+   Work8 = S9xAPUGetByte(IAPU.Address);
    S9xAPUSetByte(Work8 | IAPU.Registers.YA.B.A, IAPU.Address);
-   Work8 &= IAPU.Registers.YA.B.A;
+   Work8 = IAPU.Registers.YA.B.A - Work8;
    APUSetZN8(Work8);
    IAPU.PC += 3;
 }
@@ -796,9 +809,9 @@ void Apu4E()
 {
    // TCLR1 abs
    Absolute();
-   uint8_t Work8 = S9xAPUGetByte(IAPU.Address);
+   Work8 = S9xAPUGetByte(IAPU.Address);
    S9xAPUSetByte(Work8 & ~IAPU.Registers.YA.B.A, IAPU.Address);
-   Work8 &= IAPU.Registers.YA.B.A;
+   Work8 = IAPU.Registers.YA.B.A - Work8;
    APUSetZN8(Work8);
    IAPU.PC += 3;
 }
@@ -812,23 +825,22 @@ void Apu0F()
    Push(IAPU.Registers.P);
    APUSetBreak();
    APUClearInterrupt();
-   // XXX:Where is the BRK vector ???
-   IAPU.PC = IAPU.RAM + APU.ExtraRAM[0x20] + (APU.ExtraRAM[0x21] << 8);
+   IAPU.PC = IAPU.RAM + S9xAPUGetByte(0xffde) + (S9xAPUGetByte(0xffdf) << 8);
 }
 
 void ApuEF()
 {
    // SLEEP
-   // XXX: sleep
+   APU.TimerEnabled[0] = APU.TimerEnabled[1] = APU.TimerEnabled[2] = false;
    IAPU.APUExecuting = false;
-   IAPU.PC++;
 }
 
 void ApuFF()
 {
    // STOP
+   APU.TimerEnabled[0] = APU.TimerEnabled[1] = APU.TimerEnabled[2] = false;
    IAPU.APUExecuting = false;
-   IAPU.PC++;
+   Settings.APUEnabled = false; // re-enabled on next APU reset
 }
 
 void Apu10()
@@ -973,8 +985,7 @@ void Apu40()
 void Apu1A()
 {
    // DECW dp
-   uint16_t Work16 = S9xAPUGetByteZ(OP1) + (S9xAPUGetByteZ(OP1 + 1) << 8);
-   Work16--;
+   Work16 = S9xAPUGetByteZ(OP1) + (S9xAPUGetByteZ(OP1 + 1) << 8) - 1;
    S9xAPUSetByteZ((uint8_t) Work16, OP1);
    S9xAPUSetByteZ(Work16 >> 8, OP1 + 1);
    APUSetZN16(Work16);
@@ -984,8 +995,8 @@ void Apu1A()
 void Apu5A()
 {
    // CMPW YA,dp
-   uint16_t Work16 = S9xAPUGetByteZ(OP1) + (S9xAPUGetByteZ(OP1 + 1) << 8);
-   int32_t Int32 = (long) IAPU.Registers.YA.W - (long) Work16;
+   Work16 = S9xAPUGetByteZ(OP1) + (S9xAPUGetByteZ(OP1 + 1) << 8);
+   Int32 = (int32_t) IAPU.Registers.YA.W - (int32_t) Work16;
    IAPU._Carry = Int32 >= 0;
    APUSetZN16((uint16_t) Int32);
    IAPU.PC += 2;
@@ -994,40 +1005,36 @@ void Apu5A()
 void Apu3A()
 {
    // INCW dp
-   uint16_t Work16 = S9xAPUGetByteZ(OP1) + (S9xAPUGetByteZ(OP1 + 1) << 8);
-   Work16++;
+   Work16 = S9xAPUGetByteZ(OP1) + (S9xAPUGetByteZ(OP1 + 1) << 8) + 1;
    S9xAPUSetByteZ((uint8_t) Work16, OP1);
    S9xAPUSetByteZ(Work16 >> 8, OP1 + 1);
    APUSetZN16(Work16);
    IAPU.PC += 2;
 }
 
-// XXX: HalfCarry - BJ Fixed? Or is it between bits 7 and 8 for ADDW/SUBW?
 void Apu7A()
 {
    // ADDW YA,dp
-   uint16_t Work16 = S9xAPUGetByteZ(OP1) + (S9xAPUGetByteZ(OP1 + 1) << 8);
-   uint32_t Work32 = (uint32_t) IAPU.Registers.YA.W + Work16;
+   Work16 = S9xAPUGetByteZ(OP1) + (S9xAPUGetByteZ(OP1 + 1) << 8);
+   Work32 = (uint32_t) IAPU.Registers.YA.W + Work16;
    IAPU._Carry = Work32 >= 0x10000;
    if (~(IAPU.Registers.YA.W ^ Work16) & (Work16 ^ (uint16_t) Work32) & 0x8000)
       APUSetOverflow();
    else
       APUClearOverflow();
    APUClearHalfCarry();
-   if ((IAPU.Registers.YA.W ^ Work16 ^ (uint16_t) Work32) & 0x10)
+   if ((IAPU.Registers.YA.W ^ Work16 ^ (uint16_t) Work32) & 0x1000)
       APUSetHalfCarry();
    IAPU.Registers.YA.W = (uint16_t) Work32;
    APUSetZN16(IAPU.Registers.YA.W);
    IAPU.PC += 2;
 }
 
-// XXX: BJ: i think the old HalfCarry behavior was wrong...
-// XXX: Or is it between bits 7 and 8 for ADDW/SUBW?
 void Apu9A()
 {
    // SUBW YA,dp
-   uint16_t Work16 = S9xAPUGetByteZ(OP1) + (S9xAPUGetByteZ(OP1 + 1) << 8);
-   int32_t Int32 = (long) IAPU.Registers.YA.W - (long) Work16;
+   Work16 = S9xAPUGetByteZ(OP1) + (S9xAPUGetByteZ(OP1 + 1) << 8);
+   Int32 = (int32_t) IAPU.Registers.YA.W - (int32_t) Work16;
    APUClearHalfCarry();
    IAPU._Carry = Int32 >= 0;
    if (((IAPU.Registers.YA.W ^ Work16) & 0x8000) &&
@@ -1036,7 +1043,7 @@ void Apu9A()
    else
       APUClearOverflow();
    APUSetHalfCarry();
-   if ((IAPU.Registers.YA.W ^ Work16 ^ (uint16_t) Int32) & 0x10)
+   if ((IAPU.Registers.YA.W ^ Work16 ^ (uint16_t) Int32) & 0x1000)
       APUClearHalfCarry();
    IAPU.Registers.YA.W = (uint16_t) Int32;
    APUSetZN16(IAPU.Registers.YA.W);
@@ -1063,7 +1070,7 @@ void ApuDA()
 void Apu64()
 {
    // CMP A,dp
-   uint8_t Work8 = S9xAPUGetByteZ(OP1);
+   Work8 = S9xAPUGetByteZ(OP1);
    CMP(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC += 2;
 }
@@ -1072,7 +1079,7 @@ void Apu65()
 {
    // CMP A,abs
    Absolute();
-   uint8_t Work8 = S9xAPUGetByte(IAPU.Address);
+   Work8 = S9xAPUGetByte(IAPU.Address);
    CMP(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC += 3;
 }
@@ -1080,7 +1087,7 @@ void Apu65()
 void Apu66()
 {
    // CMP A,(X)
-   uint8_t Work8 = S9xAPUGetByteZ(IAPU.Registers.X);
+   Work8 = S9xAPUGetByteZ(IAPU.Registers.X);
    CMP(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC++;
 }
@@ -1089,7 +1096,7 @@ void Apu67()
 {
    // CMP A,(dp+X)
    IndexedXIndirect();
-   uint8_t Work8 = S9xAPUGetByte(IAPU.Address);
+   Work8 = S9xAPUGetByte(IAPU.Address);
    CMP(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC += 2;
 }
@@ -1097,7 +1104,7 @@ void Apu67()
 void Apu68()
 {
    // CMP A,#00
-   uint8_t Work8 = OP1;
+   Work8 = OP1;
    CMP(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC += 2;
 }
@@ -1105,8 +1112,8 @@ void Apu68()
 void Apu69()
 {
    // CMP dp(dest), dp(src)
-   uint8_t W1 = S9xAPUGetByteZ(OP1);
-   uint8_t Work8 = S9xAPUGetByteZ(OP2);
+   W1 = S9xAPUGetByteZ(OP1);
+   Work8 = S9xAPUGetByteZ(OP2);
    CMP(Work8, W1);
    IAPU.PC += 3;
 }
@@ -1114,7 +1121,7 @@ void Apu69()
 void Apu74()
 {
    // CMP A, dp+X
-   uint8_t Work8 = S9xAPUGetByteZ(OP1 + IAPU.Registers.X);
+   Work8 = S9xAPUGetByteZ(OP1 + IAPU.Registers.X);
    CMP(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC += 2;
 }
@@ -1123,7 +1130,7 @@ void Apu75()
 {
    // CMP A,abs+X
    AbsoluteX();
-   uint8_t Work8 = S9xAPUGetByte(IAPU.Address);
+   Work8 = S9xAPUGetByte(IAPU.Address);
    CMP(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC += 3;
 }
@@ -1132,7 +1139,7 @@ void Apu76()
 {
    // CMP A, abs+Y
    AbsoluteY();
-   uint8_t Work8 = S9xAPUGetByte(IAPU.Address);
+   Work8 = S9xAPUGetByte(IAPU.Address);
    CMP(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC += 3;
 }
@@ -1141,7 +1148,7 @@ void Apu77()
 {
    // CMP A,(dp)+Y
    IndirectIndexedY();
-   uint8_t Work8 = S9xAPUGetByte(IAPU.Address);
+   Work8 = S9xAPUGetByte(IAPU.Address);
    CMP(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC += 2;
 }
@@ -1149,8 +1156,8 @@ void Apu77()
 void Apu78()
 {
    // CMP dp,#00
-   uint8_t Work8 = OP1;
-   uint8_t W1 = S9xAPUGetByteZ(OP2);
+   Work8 = OP1;
+   W1 = S9xAPUGetByteZ(OP2);
    CMP(W1, Work8);
    IAPU.PC += 3;
 }
@@ -1158,8 +1165,8 @@ void Apu78()
 void Apu79()
 {
    // CMP (X),(Y)
-   uint8_t W1 = S9xAPUGetByteZ(IAPU.Registers.X);
-   uint8_t Work8 = S9xAPUGetByteZ(IAPU.Registers.YA.B.Y);
+   W1 = S9xAPUGetByteZ(IAPU.Registers.X);
+   Work8 = S9xAPUGetByteZ(IAPU.Registers.YA.B.Y);
    CMP(W1, Work8);
    IAPU.PC++;
 }
@@ -1168,7 +1175,7 @@ void Apu1E()
 {
    // CMP X,abs
    Absolute();
-   uint8_t Work8 = S9xAPUGetByte(IAPU.Address);
+   Work8 = S9xAPUGetByte(IAPU.Address);
    CMP(IAPU.Registers.X, Work8);
    IAPU.PC += 3;
 }
@@ -1176,7 +1183,7 @@ void Apu1E()
 void Apu3E()
 {
    // CMP X,dp
-   uint8_t Work8 = S9xAPUGetByteZ(OP1);
+   Work8 = S9xAPUGetByteZ(OP1);
    CMP(IAPU.Registers.X, Work8);
    IAPU.PC += 2;
 }
@@ -1192,7 +1199,7 @@ void Apu5E()
 {
    // CMP Y,abs
    Absolute();
-   uint8_t Work8 = S9xAPUGetByte(IAPU.Address);
+   Work8 = S9xAPUGetByte(IAPU.Address);
    CMP(IAPU.Registers.YA.B.Y, Work8);
    IAPU.PC += 3;
 }
@@ -1200,7 +1207,7 @@ void Apu5E()
 void Apu7E()
 {
    // CMP Y,dp
-   uint8_t Work8 = S9xAPUGetByteZ(OP1);
+   Work8 = S9xAPUGetByteZ(OP1);
    CMP(IAPU.Registers.YA.B.Y, Work8);
    IAPU.PC += 2;
 }
@@ -1208,7 +1215,7 @@ void Apu7E()
 void ApuAD()
 {
    // CMP Y,#00
-   uint8_t Work8 = OP1;
+   Work8 = OP1;
    CMP(IAPU.Registers.YA.B.Y, Work8);
    IAPU.PC += 2;
 }
@@ -1219,8 +1226,6 @@ void Apu1F()
    Absolute();
    IAPU.PC = IAPU.RAM + S9xAPUGetByte(IAPU.Address + IAPU.Registers.X) +
              (S9xAPUGetByte(IAPU.Address + IAPU.Registers.X + 1) << 8);
-   // XXX: HERE:
-   // APU.Flags |= TRACE_FLAG;
 }
 
 void Apu5F()
@@ -1298,7 +1303,7 @@ void Apu28()
 void Apu29()
 {
    // AND dp(dest),dp(src)
-   uint8_t Work8 = S9xAPUGetByteZ(OP1);
+   Work8 = S9xAPUGetByteZ(OP1);
    Work8 &= S9xAPUGetByteZ(OP2);
    S9xAPUSetByteZ(Work8, OP2);
    APUSetZN8(Work8);
@@ -1343,7 +1348,7 @@ void Apu37()
 void Apu38()
 {
    // AND dp,#00
-   uint8_t Work8 = OP1;
+   Work8 = OP1;
    Work8 &= S9xAPUGetByteZ(OP2);
    S9xAPUSetByteZ(Work8, OP2);
    APUSetZN8(Work8);
@@ -1353,8 +1358,7 @@ void Apu38()
 void Apu39()
 {
    // AND (X),(Y)
-   uint8_t Work8 = S9xAPUGetByteZ(IAPU.Registers.X) & S9xAPUGetByteZ(
-                    IAPU.Registers.YA.B.Y);
+   Work8 = S9xAPUGetByteZ(IAPU.Registers.X) & S9xAPUGetByteZ(IAPU.Registers.YA.B.Y);
    APUSetZN8(Work8);
    S9xAPUSetByteZ(Work8, IAPU.Registers.X);
    IAPU.PC++;
@@ -1363,7 +1367,7 @@ void Apu39()
 void Apu2B()
 {
    // ROL dp
-   uint8_t Work8 = S9xAPUGetByteZ(OP1);
+   Work8 = S9xAPUGetByteZ(OP1);
    ROL(Work8);
    S9xAPUSetByteZ(Work8, OP1);
    IAPU.PC += 2;
@@ -1373,7 +1377,7 @@ void Apu2C()
 {
    // ROL abs
    Absolute();
-   uint8_t Work8 = S9xAPUGetByte(IAPU.Address);
+   Work8 = S9xAPUGetByte(IAPU.Address);
    ROL(Work8);
    S9xAPUSetByte(Work8, IAPU.Address);
    IAPU.PC += 3;
@@ -1382,7 +1386,7 @@ void Apu2C()
 void Apu3B()
 {
    // ROL dp+X
-   uint8_t Work8 = S9xAPUGetByteZ(OP1 + IAPU.Registers.X);
+   Work8 = S9xAPUGetByteZ(OP1 + IAPU.Registers.X);
    ROL(Work8);
    S9xAPUSetByteZ(Work8, OP1 + IAPU.Registers.X);
    IAPU.PC += 2;
@@ -1398,7 +1402,7 @@ void Apu3C()
 void Apu2E()
 {
    // CBNE dp,rel
-   uint8_t Work8 = OP1;
+   Work8 = OP1;
    Relative2();
 
    if (S9xAPUGetByteZ(Work8) != IAPU.Registers.YA.B.A)
@@ -1414,7 +1418,7 @@ void Apu2E()
 void ApuDE()
 {
    // CBNE dp+X,rel
-   uint8_t Work8 = OP1 + IAPU.Registers.X;
+   Work8 = OP1 + IAPU.Registers.X;
    Relative2();
 
    if (S9xAPUGetByteZ(Work8) != IAPU.Registers.YA.B.A)
@@ -1482,7 +1486,7 @@ void ApuDC()
 void ApuAB()
 {
    // INC dp
-   uint8_t Work8 = S9xAPUGetByteZ(OP1) + 1;
+   Work8 = S9xAPUGetByteZ(OP1) + 1;
    S9xAPUSetByteZ(Work8, OP1);
    APUSetZN8(Work8);
 
@@ -1497,7 +1501,7 @@ void ApuAC()
 {
    // INC abs
    Absolute();
-   uint8_t Work8 = S9xAPUGetByte(IAPU.Address) + 1;
+   Work8 = S9xAPUGetByte(IAPU.Address) + 1;
    S9xAPUSetByte(Work8, IAPU.Address);
    APUSetZN8(Work8);
 
@@ -1511,7 +1515,7 @@ void ApuAC()
 void ApuBB()
 {
    // INC dp+X
-   uint8_t Work8 = S9xAPUGetByteZ(OP1 + IAPU.Registers.X) + 1;
+   Work8 = S9xAPUGetByteZ(OP1 + IAPU.Registers.X) + 1;
    S9xAPUSetByteZ(Work8, OP1 + IAPU.Registers.X);
    APUSetZN8(Work8);
 
@@ -1538,7 +1542,7 @@ void ApuBC()
 void Apu8B()
 {
    // DEC dp
-   uint8_t Work8 = S9xAPUGetByteZ(OP1) - 1;
+   Work8 = S9xAPUGetByteZ(OP1) - 1;
    S9xAPUSetByteZ(Work8, OP1);
    APUSetZN8(Work8);
 
@@ -1553,7 +1557,7 @@ void Apu8C()
 {
    // DEC abs
    Absolute();
-   uint8_t Work8 = S9xAPUGetByte(IAPU.Address) - 1;
+   Work8 = S9xAPUGetByte(IAPU.Address) - 1;
    S9xAPUSetByte(Work8, IAPU.Address);
    APUSetZN8(Work8);
 
@@ -1567,7 +1571,7 @@ void Apu8C()
 void Apu9B()
 {
    // DEC dp+X
-   uint8_t Work8 = S9xAPUGetByteZ(OP1 + IAPU.Registers.X) - 1;
+   Work8 = S9xAPUGetByteZ(OP1 + IAPU.Registers.X) - 1;
    S9xAPUSetByteZ(Work8, OP1 + IAPU.Registers.X);
    APUSetZN8(Work8);
 
@@ -1636,7 +1640,7 @@ void Apu48()
 void Apu49()
 {
    // EOR dp(dest),dp(src)
-   uint8_t Work8 = S9xAPUGetByteZ(OP1);
+   Work8 = S9xAPUGetByteZ(OP1);
    Work8 ^= S9xAPUGetByteZ(OP2);
    S9xAPUSetByteZ(Work8, OP2);
    APUSetZN8(Work8);
@@ -1681,7 +1685,7 @@ void Apu57()
 void Apu58()
 {
    // EOR dp,#00
-   uint8_t Work8 = OP1;
+   Work8 = OP1;
    Work8 ^= S9xAPUGetByteZ(OP2);
    S9xAPUSetByteZ(Work8, OP2);
    APUSetZN8(Work8);
@@ -1691,8 +1695,7 @@ void Apu58()
 void Apu59()
 {
    // EOR (X),(Y)
-   uint8_t Work8 = S9xAPUGetByteZ(IAPU.Registers.X) ^ S9xAPUGetByteZ(
-                    IAPU.Registers.YA.B.Y);
+   Work8 = S9xAPUGetByteZ(IAPU.Registers.X) ^ S9xAPUGetByteZ(IAPU.Registers.YA.B.Y);
    APUSetZN8(Work8);
    S9xAPUSetByteZ(Work8, IAPU.Registers.X);
    IAPU.PC++;
@@ -1701,7 +1704,7 @@ void Apu59()
 void Apu4B()
 {
    // LSR dp
-   uint8_t Work8 = S9xAPUGetByteZ(OP1);
+   Work8 = S9xAPUGetByteZ(OP1);
    LSR(Work8);
    S9xAPUSetByteZ(Work8, OP1);
    IAPU.PC += 2;
@@ -1711,7 +1714,7 @@ void Apu4C()
 {
    // LSR abs
    Absolute();
-   uint8_t Work8 = S9xAPUGetByte(IAPU.Address);
+   Work8 = S9xAPUGetByte(IAPU.Address);
    LSR(Work8);
    S9xAPUSetByte(Work8, IAPU.Address);
    IAPU.PC += 3;
@@ -1720,7 +1723,7 @@ void Apu4C()
 void Apu5B()
 {
    // LSR dp+X
-   uint8_t Work8 = S9xAPUGetByteZ(OP1 + IAPU.Registers.X);
+   Work8 = S9xAPUGetByteZ(OP1 + IAPU.Registers.X);
    LSR(Work8);
    S9xAPUSetByteZ(Work8, OP1 + IAPU.Registers.X);
    IAPU.PC += 2;
@@ -1783,7 +1786,7 @@ void ApuBD()
 void Apu6B()
 {
    // ROR dp
-   uint8_t Work8 = S9xAPUGetByteZ(OP1);
+   Work8 = S9xAPUGetByteZ(OP1);
    ROR(Work8);
    S9xAPUSetByteZ(Work8, OP1);
    IAPU.PC += 2;
@@ -1793,7 +1796,7 @@ void Apu6C()
 {
    // ROR abs
    Absolute();
-   uint8_t Work8 = S9xAPUGetByte(IAPU.Address);
+   Work8 = S9xAPUGetByte(IAPU.Address);
    ROR(Work8);
    S9xAPUSetByte(Work8, IAPU.Address);
    IAPU.PC += 3;
@@ -1802,7 +1805,7 @@ void Apu6C()
 void Apu7B()
 {
    // ROR dp+X
-   uint8_t Work8 = S9xAPUGetByteZ(OP1 + IAPU.Registers.X);
+   Work8 = S9xAPUGetByteZ(OP1 + IAPU.Registers.X);
    ROR(Work8);
    S9xAPUSetByteZ(Work8, OP1 + IAPU.Registers.X);
    IAPU.PC += 2;
@@ -1818,9 +1821,9 @@ void Apu7C()
 void Apu6E()
 {
    // DBNZ dp,rel
-   uint8_t Work8 = OP1;
+   Work8 = OP1;
    Relative2();
-   uint8_t W1 = S9xAPUGetByteZ(Work8) - 1;
+   W1 = S9xAPUGetByteZ(Work8) - 1;
    S9xAPUSetByteZ(W1, Work8);
    if (W1 != 0)
    {
@@ -1864,7 +1867,7 @@ void Apu7F()
 void Apu84()
 {
    // ADC A,dp
-   uint8_t Work8 = S9xAPUGetByteZ(OP1);
+   Work8 = S9xAPUGetByteZ(OP1);
    ADC(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC += 2;
 }
@@ -1873,7 +1876,7 @@ void Apu85()
 {
    // ADC A, abs
    Absolute();
-   uint8_t Work8 = S9xAPUGetByte(IAPU.Address);
+   Work8 = S9xAPUGetByte(IAPU.Address);
    ADC(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC += 3;
 }
@@ -1881,7 +1884,7 @@ void Apu85()
 void Apu86()
 {
    // ADC A,(X)
-   uint8_t Work8 = S9xAPUGetByteZ(IAPU.Registers.X);
+   Work8 = S9xAPUGetByteZ(IAPU.Registers.X);
    ADC(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC++;
 }
@@ -1890,7 +1893,7 @@ void Apu87()
 {
    // ADC A,(dp+X)
    IndexedXIndirect();
-   uint8_t Work8 = S9xAPUGetByte(IAPU.Address);
+   Work8 = S9xAPUGetByte(IAPU.Address);
    ADC(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC += 2;
 }
@@ -1898,7 +1901,7 @@ void Apu87()
 void Apu88()
 {
    // ADC A,#00
-   uint8_t Work8 = OP1;
+   Work8 = OP1;
    ADC(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC += 2;
 }
@@ -1906,8 +1909,8 @@ void Apu88()
 void Apu89()
 {
    // ADC dp(dest),dp(src)
-   uint8_t Work8 = S9xAPUGetByteZ(OP1);
-   uint8_t W1 = S9xAPUGetByteZ(OP2);
+   Work8 = S9xAPUGetByteZ(OP1);
+   W1 = S9xAPUGetByteZ(OP2);
    ADC(W1, Work8);
    S9xAPUSetByteZ(W1, OP2);
    IAPU.PC += 3;
@@ -1916,7 +1919,7 @@ void Apu89()
 void Apu94()
 {
    // ADC A,dp+X
-   uint8_t Work8 = S9xAPUGetByteZ(OP1 + IAPU.Registers.X);
+   Work8 = S9xAPUGetByteZ(OP1 + IAPU.Registers.X);
    ADC(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC += 2;
 }
@@ -1925,7 +1928,7 @@ void Apu95()
 {
    // ADC A, abs+X
    AbsoluteX();
-   uint8_t Work8 = S9xAPUGetByte(IAPU.Address);
+   Work8 = S9xAPUGetByte(IAPU.Address);
    ADC(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC += 3;
 }
@@ -1934,7 +1937,7 @@ void Apu96()
 {
    // ADC A, abs+Y
    AbsoluteY();
-   uint8_t Work8 = S9xAPUGetByte(IAPU.Address);
+   Work8 = S9xAPUGetByte(IAPU.Address);
    ADC(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC += 3;
 }
@@ -1943,7 +1946,7 @@ void Apu97()
 {
    // ADC A, (dp)+Y
    IndirectIndexedY();
-   uint8_t Work8 = S9xAPUGetByte(IAPU.Address);
+   Work8 = S9xAPUGetByte(IAPU.Address);
    ADC(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC += 2;
 }
@@ -1951,8 +1954,8 @@ void Apu97()
 void Apu98()
 {
    // ADC dp,#00
-   uint8_t Work8 = OP1;
-   uint8_t W1 = S9xAPUGetByteZ(OP2);
+   Work8 = OP1;
+   W1 = S9xAPUGetByteZ(OP2);
    ADC(W1, Work8);
    S9xAPUSetByteZ(W1, OP2);
    IAPU.PC += 3;
@@ -1961,8 +1964,8 @@ void Apu98()
 void Apu99()
 {
    // ADC (X),(Y)
-   uint8_t W1 = S9xAPUGetByteZ(IAPU.Registers.X);
-   uint8_t Work8 = S9xAPUGetByteZ(IAPU.Registers.YA.B.Y);
+   W1 = S9xAPUGetByteZ(IAPU.Registers.X);
+   Work8 = S9xAPUGetByteZ(IAPU.Registers.YA.B.Y);
    ADC(W1, Work8);
    S9xAPUSetByteZ(W1, IAPU.Registers.X);
    IAPU.PC++;
@@ -1979,7 +1982,7 @@ void Apu8D()
 void Apu8F()
 {
    // MOV dp,#00
-   uint8_t Work8 = OP1;
+   Work8 = OP1;
    S9xAPUSetByteZ(Work8, OP2);
    IAPU.PC += 3;
 }
@@ -1987,21 +1990,36 @@ void Apu8F()
 void Apu9E()
 {
    // DIV YA,X
-   if (IAPU.Registers.X == 0)
-   {
-      APUSetOverflow();
-      IAPU.Registers.YA.B.Y = 0xff;
-      IAPU.Registers.YA.B.A = 0xff;
-   }
+   if ((IAPU.Registers.X & 0x0f) <= (IAPU.Registers.YA.B.Y & 0x0f))
+      APUSetHalfCarry();
    else
+      APUClearHalfCarry();
+
+   uint32_t yva = IAPU.Registers.YA.W;
+   uint32_t x = IAPU.Registers.X << 9;
+   uint32_t i;
+
+   for (i = 0 ; i < 9 ; ++i)
    {
-      APUClearOverflow();
-      uint8_t Work8 = IAPU.Registers.YA.W / IAPU.Registers.X;
-      IAPU.Registers.YA.B.Y = IAPU.Registers.YA.W % IAPU.Registers.X;
-      IAPU.Registers.YA.B.A = Work8;
+      yva <<= 1;
+
+      if (yva & 0x20000)
+         yva = (yva & 0x1ffff) | 1;
+
+      if (yva >= x)
+         yva ^= 1;
+      
+      if (yva & 1)
+         yva = (yva - x) & 0x1ffff;
    }
-   // XXX How should Overflow, Half Carry, Zero and Negative flags be set??
-   // APUSetZN16 (IAPU.Registers.YA.W);
+
+   if (yva & 0x100)
+       APUSetOverflow();
+   else
+       APUClearOverflow();
+
+   IAPU.Registers.YA.B.Y = (yva >> 9) & 0xff;
+   IAPU.Registers.YA.B.A = yva & 0xff;
    APUSetZN8(IAPU.Registers.YA.B.A);
    IAPU.PC++;
 }
@@ -2009,8 +2027,7 @@ void Apu9E()
 void Apu9F()
 {
    // XCN A
-   IAPU.Registers.YA.B.A = (IAPU.Registers.YA.B.A >> 4) | (IAPU.Registers.YA.B.A <<
-                           4);
+   IAPU.Registers.YA.B.A = (IAPU.Registers.YA.B.A >> 4) | (IAPU.Registers.YA.B.A << 4);
    APUSetZN8(IAPU.Registers.YA.B.A);
    IAPU.PC++;
 }
@@ -2018,7 +2035,7 @@ void Apu9F()
 void ApuA4()
 {
    // SBC A, dp
-   uint8_t Work8 = S9xAPUGetByteZ(OP1);
+   Work8 = S9xAPUGetByteZ(OP1);
    SBC(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC += 2;
 }
@@ -2027,7 +2044,7 @@ void ApuA5()
 {
    // SBC A, abs
    Absolute();
-   uint8_t Work8 = S9xAPUGetByte(IAPU.Address);
+   Work8 = S9xAPUGetByte(IAPU.Address);
    SBC(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC += 3;
 }
@@ -2035,7 +2052,7 @@ void ApuA5()
 void ApuA6()
 {
    // SBC A, (X)
-   uint8_t Work8 = S9xAPUGetByteZ(IAPU.Registers.X);
+   Work8 = S9xAPUGetByteZ(IAPU.Registers.X);
    SBC(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC++;
 }
@@ -2044,7 +2061,7 @@ void ApuA7()
 {
    // SBC A,(dp+X)
    IndexedXIndirect();
-   uint8_t Work8 = S9xAPUGetByte(IAPU.Address);
+   Work8 = S9xAPUGetByte(IAPU.Address);
    SBC(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC += 2;
 }
@@ -2052,7 +2069,7 @@ void ApuA7()
 void ApuA8()
 {
    // SBC A,#00
-   uint8_t Work8 = OP1;
+   Work8 = OP1;
    SBC(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC += 2;
 }
@@ -2060,8 +2077,8 @@ void ApuA8()
 void ApuA9()
 {
    // SBC dp(dest), dp(src)
-   uint8_t Work8 = S9xAPUGetByteZ(OP1);
-   uint8_t W1 = S9xAPUGetByteZ(OP2);
+   Work8 = S9xAPUGetByteZ(OP1);
+   W1 = S9xAPUGetByteZ(OP2);
    SBC(W1, Work8);
    S9xAPUSetByteZ(W1, OP2);
    IAPU.PC += 3;
@@ -2070,7 +2087,7 @@ void ApuA9()
 void ApuB4()
 {
    // SBC A, dp+X
-   uint8_t Work8 = S9xAPUGetByteZ(OP1 + IAPU.Registers.X);
+   Work8 = S9xAPUGetByteZ(OP1 + IAPU.Registers.X);
    SBC(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC += 2;
 }
@@ -2079,7 +2096,7 @@ void ApuB5()
 {
    // SBC A,abs+X
    AbsoluteX();
-   uint8_t Work8 = S9xAPUGetByte(IAPU.Address);
+   Work8 = S9xAPUGetByte(IAPU.Address);
    SBC(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC += 3;
 }
@@ -2088,7 +2105,7 @@ void ApuB6()
 {
    // SBC A,abs+Y
    AbsoluteY();
-   uint8_t Work8 = S9xAPUGetByte(IAPU.Address);
+   Work8 = S9xAPUGetByte(IAPU.Address);
    SBC(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC += 3;
 }
@@ -2097,7 +2114,7 @@ void ApuB7()
 {
    // SBC A,(dp)+Y
    IndirectIndexedY();
-   uint8_t Work8 = S9xAPUGetByte(IAPU.Address);
+   Work8 = S9xAPUGetByte(IAPU.Address);
    SBC(IAPU.Registers.YA.B.A, Work8);
    IAPU.PC += 2;
 }
@@ -2105,8 +2122,8 @@ void ApuB7()
 void ApuB8()
 {
    // SBC dp,#00
-   uint8_t Work8 = OP1;
-   uint8_t W1 = S9xAPUGetByteZ(OP2);
+   Work8 = OP1;
+   W1 = S9xAPUGetByteZ(OP2);
    SBC(W1, Work8);
    S9xAPUSetByteZ(W1, OP2);
    IAPU.PC += 3;
@@ -2115,8 +2132,8 @@ void ApuB8()
 void ApuB9()
 {
    // SBC (X),(Y)
-   uint8_t W1 = S9xAPUGetByteZ(IAPU.Registers.X);
-   uint8_t Work8 = S9xAPUGetByteZ(IAPU.Registers.YA.B.Y);
+   W1 = S9xAPUGetByteZ(IAPU.Registers.X);
+   Work8 = S9xAPUGetByteZ(IAPU.Registers.YA.B.Y);
    SBC(W1, Work8);
    S9xAPUSetByteZ(W1, IAPU.Registers.X);
    IAPU.PC++;
@@ -2132,15 +2149,17 @@ void ApuAF()
 void ApuBE()
 {
    // DAS
-   if ((IAPU.Registers.YA.B.A & 0x0f) > 9 || !APUCheckHalfCarry())
-      IAPU.Registers.YA.B.A -= 6;
-   if (IAPU.Registers.YA.B.A > 0x9f || !IAPU._Carry)
+   if (IAPU.Registers.YA.B.A > 0x99 || !IAPU._Carry)
    {
       IAPU.Registers.YA.B.A -= 0x60;
       APUClearCarry();
    }
    else
       APUSetCarry();
+
+   if ((IAPU.Registers.YA.B.A & 0x0f) > 9 || !APUCheckHalfCarry())
+      IAPU.Registers.YA.B.A -= 6;
+
    APUSetZN8(IAPU.Registers.YA.B.A);
    IAPU.PC++;
 }
@@ -2232,7 +2251,7 @@ void ApuCF()
 {
    // MUL YA
    IAPU.Registers.YA.W = (uint16_t) IAPU.Registers.YA.B.A * IAPU.Registers.YA.B.Y;
-   APUSetZN16(IAPU.Registers.YA.W);
+   APUSetZN8(IAPU.Registers.YA.B.Y);
    IAPU.PC++;
 }
 
@@ -2291,20 +2310,17 @@ void ApuDB()
 void ApuDF()
 {
    // DAA
-   if ((IAPU.Registers.YA.B.A & 0x0f) > 9 || APUCheckHalfCarry())
-   {
-      if (IAPU.Registers.YA.B.A > 0xf0) APUSetCarry();
-      IAPU.Registers.YA.B.A += 6;
-      //APUSetHalfCarry (); Intel procs do this, but this is a Sony proc...
-   }
-   //else { APUClearHalfCarry (); } ditto as above
-   if (IAPU.Registers.YA.B.A > 0x9f || IAPU._Carry)
+   if (IAPU.Registers.YA.B.A > 0x99 || IAPU._Carry)
    {
       IAPU.Registers.YA.B.A += 0x60;
       APUSetCarry();
    }
    else
       APUClearCarry();
+
+   if ((IAPU.Registers.YA.B.A & 0x0f) > 9 || APUCheckHalfCarry())
+      IAPU.Registers.YA.B.A += 6;
+
    APUSetZN8(IAPU.Registers.YA.B.A);
    IAPU.PC++;
 }